J. Hoffmann and S. Edelkamp

Volume 24, 2005

**Links to Full Text:**

We provide an overview of the organization and results of the deterministic part of the 4th International Planning Competition, i.e., of the part concerned with evaluating systems doing deterministic planning. IPC-4 attracted even more competing systems than its already large predecessors, and the competition event was revised in several important respects. After giving an introduction to the IPC, we briefly explain the main differences between the deterministic part of IPC-4 and its predecessors. We then introduce formally the language used, called PDDL2.2 that extends PDDL2.1 by derived predicates and timed initial literals. We list the competing systems and overview the results of the competition. The entire set of data is far too large to be presented in full. We provide a detailed summary; the complete data is available in an online appendix. We explain how we awarded the competition prizes.

Journal of Artificial Intelligence Research 24 (2005) 519-579 Submitted 12/04; published 10/05 The Deterministic Part of IPC-4: An Overview J"org Hoffmann hoffmann@mpi-sb.mpg.de Max-Planck-Institut f"ur Informatik, Saarbr"ucken, Germany Stefan Edelkamp stefan.edelkamp@cs.uni-dortmund.de Fachbereich Informatik, Universit"at Dortmund, Germany Abstract We provide an overview of the organization and results of the deterministic part ofthe 4th International Planning Competition, i.e., of the part concerned with evaluating systems doing deterministic planning. IPC-4 attracted even more competing systems thanits already large predecessors, and the competition event was revised in several important respects. After giving an introduction to the IPC, we briefly explain the main differencesbetween the deterministic part of IPC-4 and its predecessors. We then introduce formally the language used, called PDDL2.2 that extends PDDL2.1 by derived predicates and timedinitial literals. We list the competing systems and overview the results of the competition. The entire set of data is far too large to be presented in full. We provide a detailed summary;the complete data is available in an online appendix. We explain how we awarded the competition prizes. 1. Introduction In the application of Artificial Intelligence technology to the real-world, time and space resources are usually limited. This has led to a performance-oriented interpretation of AI in many of its research branches. Competition events have been established in automated theorem proving, in satisfiability testing, and, in particular, in AI Planning. A competition provides a large-scale evaluation platform. Due to the broadness and neutrality of that platform, a competition is far better at assessing the state-of-the-art in a research branch than the experiments ran by individual authors: more systems are compared, and the benchmarks are chosen by the competition organizers, rather than by the system authors themselves. Moreover, a competition can serve to establish a common representation formalism, and a common core set of benchmarks, marking the edge of current system capabilities. The International Planning Competition (IPC) is a biennial event, hosted at the international conferences on AI Planning and Scheduling. The IPC began in 1998 when Drew McDermott and a committee created a common specification language (PDDL) and a collection of problems forming a first benchmark (McDermott et al., 1998). PDDL is a Lisp-like input language description format that includes planning formalisms like STRIPS (Fikes & Nilsson, 1971). Five systems participated in the first international planning competition, IPC-1 for short, hosted at AIPS 1998 in Pittsburgh, Pennsylvania (McDermott, 2000). In the year 2000, Fahiem Bacchus continued this work, and the IPC-2 event attracted 16 competitors (Bacchus, 2001). The event - hosted at AIPS 2000 in Breckenridge, Colorado - was extended to include both fully automatic and hand-tailored planning systems. As cfl2005 AI Access Foundation. All rights reserved. Hoffmann & Edelkamp hand-tailored planners were allowed to use some additional domain-dependent information to the PDDL input in order to improve their performance, they participated in an additional, separate, track. Both STRIPS and ADL (Pednault, 1989) domains were used but no further extensions were made to the language (Bacchus, 2000). The 3rd International Planning Competition, IPC-3, was run by Derek Long and Maria Fox and was hosted at AIPS 2002, Toulouse, France. The competition attracted 14 competitors (Long & Fox, 2003), and focussed on planning in temporal and metric domains. For that purpose, Fox and Long developed the PDDL2.1 language (Fox & Long, 2003), of which the first three levels were used in IPC-3. Level 1 was STRIPS and ADL planning as before, Level 2 added numeric variables, Level 3 added durational constructs. The 4th International Planning Competition, IPC-4, was hosted at ICAPS-2004, Whistler, Canada. IPC-4 built on the previous efforts, in particular the language PDDL2.1. The competition event was extended and revised in several respects. In particular, IPC-4 featured, for the first time, a competition for probabilistic planners, so that the overall competition was split into a deterministic part - a continuation of the previous events - as well as a probabilistic part.1 In the latter part, co-organized by Michael Littman and H*akan Younes, the main objective of the event was to introduce a common representation language for probabilistic planners, and to establish some first benchmarks and results. For more information on the probabilistic part of IPC-4 see the work of Younes and Littman (2005). Herein, we provide an overview of the organization and results of the deterministic part of IPC-4. With 19 competing systems (21 when counting different system versions), the event was even a little larger than its already large predecessors. Several important revisions were made to the event. We briefly explain the main differences in Section 2. Afterwards, Section 3 describes the input language used, named PDDL2.2: the first three levels of PDDL2.1, extended with derived predicates and timed initial literals. Section 4 lists and briefly explains the competing systems. Section 5 then presents, for each benchmark domain, a selection of results plots, highlighting the most important points. The entire set of data points is far too large to be presented in detail. The full data, including plots for all results, is available in an online appendix. Section 6 explains how we awarded the competition prizes, Section 7 closes the paper with some concluding remarks. Appendix A gives a BNF description of PDDL2.2. 2. Main Revisions made in IPC-4 The main revisions we made to the deterministic part of IPC-4, in difference to its predecessors, were the following. Competition Workshop. We ran an international workshop on the competition one year before the event itself (Edelkamp & Hoffmann, 2003), providing the involved groups of people (system developers, organizing committee, AI Planning researchers in general) with an opportunity to express their views on issues related to the IPC. Discussions, technical talks, and panels covered all relevant topics ranging from the event's organizational structure and its input language to the selection of benchmarks and the evaluation of results. The 1. While IPC-4 was running, the deterministic part was named "classical" part. We re-named it into "deterministic" part since that wording is less ambiguous. 520 The Deterministic Part of IPC-4: An Overview workshop was especially useful for us as the organizers, giving us direct feedback on the event's organization. PDDL Extensions. There was large agreement in the community that PDDL2.1 still posed a significant challenge, and so the IPC-4 language featured only relatively minor extensions. We added language features for derived predicates and timed initial literals. The resulting language is called PDDL2.2, and keeps the 3-leveled structure of PDDL2.1. Both new language features are practically motivated and were put to use in some of the IPC-4 domains. Derived predicates add a form of domain axioms to PDDL2.1. A typical use is to formulate indirect consequences of a planner's actions. Timed initial literals add to the temporal part of PDDL2.1 (level 3) a way of defining literals that will become true at a certain time point, independently of the actions taken by the planner. A typical use is to formulate time windows and/or goal deadlines. Application-Oriented Benchmarks. Our main effort in the organization of IPC-4 was to devise a range of interesting benchmark domains, oriented at (and as close as possible to) real-world application domains. We collaborated with a number of people to achieve this goal. The description of the application domains, and of our PDDL2.2 adaptations, is long (50+ pages). It is submitted to this same JAIR special track (Edelkamp, Hoffmann, Englert, Liporace, Thiebaux, & Tr"ug, 2005). The application domains we modelled were: * Airport, modelling airport ground traffic control (Hatzack & Nebel, 2001; Tr"ug, Hoffmann, & Nebel, 2004). * Pipesworld, modelling oil derivative transportation in pipeline networks (Milidiu, dos Santos Liporace, & de Lucena, 2003; Milidi'u & dos Santos Liporace, 2004). * Promela, modelling deadlock detection in communication protocols formulated in the Promela language (Edelkamp, 2003b, 2003a). * PSR, a deterministic variant of the Power Supply Restoration benchmark (Bertoli, Cimatti, Slaney, & Thi'ebaux, 2002; Bonet & Thiebaux, 2003). * UMTS, modelling the task of scheduling the concurrent application setup in UMTS mobile devices (Englert, 2003, 2005). In addition, we re-used the Satellite and Settlers domains from IPC-3. In the case of Satellite, we added some additional domain versions to model more advanced aspects of the application (namely, the sending of the data to earth). Each domain is described in a little more detail in the part (sub-section) of Section 5 that contains the respective competition results. Domain Compilations. For the first time in the IPC, we provided domain formulations where problem constraints were compiled from the more complex PDDL subsets into the less complex subsets. In many of our domains the most natural domain formulation comes with complex precondition formulas and conditional effects, i.e., in ADL. We compiled these domain formulations to STRIPS. In previous IPCs, for example in the Elevator domain used in IPC-2, the more interesting problem constraints were dropped in the STRIPS domain versions. By using the compilation approach, we hope and believe to have created a structurally much more interesting range of STRIPS benchmarks. The ADL 521 Hoffmann & Edelkamp and STRIPS encodings of the same instances were posed to the competitors in an optional way, i.e., of every domain version there could be several domain version formulations. The competitors were allowed to choose the formulation they liked best. The results within a domain version were evaluated together, in order to keep the number of separation lines in the data at an acceptable level. We also employed compilations encoding the new PDDL2.2 language features in terms of artificial constructs with the old PDDL2.1 features. By this we intended to enable as wide participation in the domains as possible. The compiled domains were offered as separate domain versions rather than alternative formulations because, in difference to the ADL/STRIPS case, we figured that the compiled domains were too different from the original ones to allow joint evaluation. Most importantly, when compiling derived predicates or timed initial literals away, the plan length increases. Details about the compilation methods and about the arrangement of the individual domains are in the paper describing these domains (Edelkamp et al., 2005). Optimal vs. Satisficing Planning. We define optimal planners as planners that prove a guarantee on the quality of the found solution. Opposed to that, satisficing planners are planners that do not prove any guarantee other than the correctness of the solution.2 Previous IPCs did not distinguish between optimal planners and satisficing planners. But that is not fair since optimal planners are essentially solving a different problem. The theoretical hardness of a domain can differ for the different kinds of planners. In fact, it was recently proved that, in most benchmark domains, satisficing planning is easy (polynomial) while optimal planning is hard (NP-complete) (Helmert, 2001, 2003). In practice, i.e., in most of the commonly used benchmark domains, nowadays there is indeed a huge performance gap between optimal and satisficing planners. In IPC-4, we separated them into different tracks. The optimal track attracted seven systems; for example, the planning as satisfiability approach, that had disappeared in IPC-3, resurfaced. Competition Booklet and Results Posters. At previous competitions, at conference time one could neither access the results obtained in the competition, nor descriptions of the competing planners. This is clearly a drawback, especially given the growing complexity of the event. For ICAPS 2004, we assembled a booklet containing extended abstracts describing the core aspects of all competing systems (Edelkamp, Hoffmann, Littman, & Younes, 2004); the booklet was distributed to all conference participants. The competition results, i.e., those of the deterministic part, were made available in the form of posters showing runtime and plan quality plots. IPC Web-page. An important repository for any large-scale competition event is a web page containing all the relevant information - benchmark problems, language descriptions, result files, etc. For IPC-4, we have set this page up at the permanent Internet address http://ipc.icaps-conference.org. In the long run, this address is intended to provide an entry point to the IPC event as a whole, thereby avoiding the need to look up the pages for the different IPC editions at completely separate points in the web. As a less positive change from IPCs 2 and 3 to IPC-4, the track for hand-tailored planners disappeared. The reason for this is simply that no such systems registered as competitors. 2. Satisficing planners were referred to as "sub-optimal" when IPC-4 was running. We decided to replace the term since it is a bit misleading. While not guaranteeing optimal solutions, a satisficing planner may well produce such solutions in some cases. 522 The Deterministic Part of IPC-4: An Overview Of course, this is not a coincidence. There is a large agreement in the community that there are several problems with the hand-tailored track as ran in IPC-2 and IPC-3. The most important criticism is that, for hand-tailored planners, "performance" is not just runtime and plan quality results on a set of benchmarks. What's also important - maybe the most important aspect of all - is how much effort was spent in achieving these results: how hard is it to tailor the planner? While the latter obviously is an important question, likewise obviously there isn't an easy answer. We discussed the matter with a lot of people (e.g., in the ICAPS'03 workshop), but no-one could come up with an idea that seemed adequate and feasible. So basically we offered the hand-tailored planners the opportunity to participate in a track similar to the previous ones, maybe with some additional ad-hoc measurements such as how many person hours were spent in tailoring the planner to a domain. Apart from the evaluation shortcomings of such an approach, another important reason why no hand-tailored planners registered is that participating in that track is a huge amount of work - maintaining/developing the planner plus understanding the domains and tailoring the planner. Understanding the domains would have been particularly hard with the more complex domains used in IPC-4. It is thus, obviously, difficult to find enough time to participate in a hand-tailored IPC. Some more on the future of the hand-tailored track is said in Section 7. 3. PDDL 2.2 As said, the IPC-4 competition language PDDL2.2 is an extension of the first three levels of PDDL2.1 (Fox & Long, 2003). PDDL2.2 inherits the separation into the levels. The language features added on top of PDDL2.1 are derived predicates (into levels 1,2, and 3) and timed initial literals (into level 3 only). We now discuss these two features in that order, describing their syntax and semantics. A full BNF description of PDDL2.2 is in Appendix A. 3.1 Derived Predicates Derived predicates have been implemented in several planning systems in the past, for example in UCPOP (Penberthy & Weld, 1992). They are predicates that are not affected by any of the actions available to the planner. Instead, the predicate's truth values are derived by a set of rules of the form if OE(x) then P (x). The semantics, roughly, are that an instance of a derived predicate (a derived predicate whose arguments are instantiated with constants; a fact, for short) is TRUE if and only if it can be derived using the available rules (more details below). Under the name "axioms", derived predicates were a part of the original PDDL language defined by McDermott (McDermott et al., 1998) for the first planning competition, but they have never been put to use in a competition benchmark (we use the name "derived predicates" instead of "axioms" in order to avoid confusion with safety conditions). Derived predicates combine several key aspects that made them a useful language extension for IPC-4: * They are practically motivated: in particular, they provide a concise and convenient means to express updates on the transitive closure of a relation. Such updates occur 523 Hoffmann & Edelkamp in domains that include structures such as paths or flows (electricity flows, chemical flows, etc.); in particular, the PSR domain includes this kind of structure. * They are also theoretically justified in that compiling them away can be infeasible. It was recently proved that, in the worst case, compiling derived predicates away results in an exponential blow up of either the problem description or the plan length (Thiebaux, Hoffmann, & Nebel, 2003, 2005). * Derived predicates do not cause a significant implementation overhead in, at least, the state transition routines used by forward search planners. When the world state - the truth values of all non-derived, basic, predicates - is known, computing the truth values of the derived predicates is trivial.3 In the IPC-4 benchmarks, derived predicates were used only in the non-durational context, PDDL2.2 level 1. 3.1.1 Syntax The BNF definition of derived predicates involves just two small modifications to the BNF definition of PDDL2.1: hstructure-defi ::=:derived-predicates hderived-defi The domain file specifies a list of "structures". In PDDL2.1 these were either actions or durational actions. Now we also allow "derived" definitions at these points. hderived-defi ::= (:derived hatomic formula(term)i hGDi) The "derived" definitions are the "rules" mentioned above. They simply specify the predicate P to be derived (with variable vector x), and the formula OE(x) from which instances of P can be concluded to be true. Syntactically, the predicate and variables are given by the hatomic formula(term)i expression, and the formula is given by hGDi (a "goal description", i.e. a formula). The BNF is more generous than what we actually allow in PDDL2.2, respectively in IPC-4. We make a number of restrictions to ensure that the definitions make sense and are easy to treat algorithmically. We call a predicate P derived if there is a rule that has a predicate P in its head; otherwise we call P basic. The restrictions we make are the following. 1. The actions available to the planner do not affect the derived predicates: no derived predicate occurs on any of the effect lists of the domain actions. 2. If a rule defines that P (x) can be derived from OE(x), then the variables in x are pairwise different (and, as the notation suggests, the free variables of OE(x) are exactly the variables in x). 3. Note, though, that it may be much less trivial to adapt a heuristic function to handle derived predicates; this is, for example, discussed by Thiebaux et al. (2003, 2005). 524 The Deterministic Part of IPC-4: An Overview 3. If a rule defines that P (x) can be derived from OE, then the Negation Normal Form (NNF) of OE(x) does not contain any derived predicates in negated form. The first restriction ensures that there is a separation between the predicates that the planner can affect (the basic predicates) and those (the derived predicates) whose truth values follow from the basic predicates. The second restriction ensures that the rule right hand sides match the rule left hand sides. Let us explain the third restriction. The NNF of a formula is obtained by "pushing the negations downwards", i.e. transforming ~8x : OE into9 x : (~OE), ~9x : OE into 8x : (~OE), ~ W OEi into V(~OEi), and ~ V OEi into W(~OEi). Iterating these transformation steps, one ends up with a formula where negations occur only in front of atomic formulas - predicates with variable vectors, in our case. The formula contains a predicate P in negated form if and only if there is an occurrence of P that is negated. By requiring that the formulas in the rules (that derive predicate values) do not contain any derived predicates in negated form, we ensure that there can not be any negative interactions between applications of the rules (see the semantics below). An example of a derived predicate is the "above" predicate in the Blocksworld, which is true between blocks x and y whenever x is transitively (possibly with some blocks in between) on y. Using the derived predicates syntax, this predicate can be defined as follows. (:derived (above ?x ?y) (or (on ?x ?y) (exists (?z) (and (on ?x ?z) (above ?z ?y))))) Note that formulating the truth value of "above" in terms of the effects of the normal Blocksworld actions is very awkward. Since the set of atoms affected by an action depends on the situation, one either needs artificial actions or complex conditional effects. 3.1.2 Semantics We now describe the updates that need to be made to the PDDL2.1 semantics definitions given by Fox and Long (2003). We introduce formal notations to capture the semantics of derived predicates. We then "hook" these semantics into the PDDL2.1 language by modifying two of Fox and Long (2003)'s definitions. Say we are given the truth values of all (instances of the) basic predicates, and want to compute the truth values of the (instances of the) derived predicates from that. We are in this situation every time we have applied an action, or parallel action set. In the durational context, we are in this situation at the "happenings" in our current plan, that is every time a durative action starts or finishes. Formally, what we want to have is a function D that maps a set of basic facts (instances of basic predicates) to the same set but enriched with derived facts (the derivable instances of the derived predicates). Assume we are given the set R of rules for the derived predicates, where the elements of R have the form (P (x), OE(x)) - if OE(x) then P (x). Then D(s), for a set of basic facts s, is defined as follows. D(s) := "{s0 | s ` s0, 8(P (x), OE(x)) 2 R : 8c, |c| = |x| : (s0 |= OE(c) ) P (c) 2 s0)} (1) This definition uses the standard notations of the modelling relation |= between states (represented as sets of facts in our case) and formulas, and of the substitution OE(c) of the 525 Hoffmann & Edelkamp free variables in formula OE(x) with a constant vector c. In words, D(s) is the intersection of all supersets of s that are closed under application of the rules R. Remember that we restrict the rules to not contain any derived predicates in negated form. This implies that the order in which the rules are applied to a state does not matter (we can not "lose" any derived facts by deriving other facts first). This, in turn, implies that D(s) is itself closed under application of the rules R. In other words, D(s) is the least fixed point over the possible applications of the rules R to the state where all derived facts are assumed to be FALSE (represented by their not being contained in s). More constructively, D(s) can be computed by the following simple process. s0 := s do select a rule (P (x), OE(x)) and a vector c of constants, with |c| = |x|, such that s0 |= OE(c) and P (c) 62 s0 let s0 := s0 [ {P (c)} until no such rule and constant vector exist let D(s) := s0 In words, apply the applicable rules in an arbitrary order until no new facts can be derived anymore. We can now specify what an executable plan is in PDDL2.1 with derived predicates. All we need to do is to hook the function D into Definition 13, "Happening Execution", given by Fox and Long (2003). By this definition, Fox and Long (2003) define the state transitions in a plan. The happenings in a (temporal or non-temporal) plan are all time points at which at least one action effect occurs. Fox and Long (2003)'s definition is this: Definition 13 Happening Execution (Fox & Long, 2003) Given a state, (t, s, x) and a happening, H, the activity for H is the set of grounded actions AH = {a| the name for a is in H, a is valid and P rea is satisfied in (t, s, x)} The result of executing a happening, H, associated with time tH, in a state (t, s, x) is undefined if |AH | 6= |H| or if any pair of actions in AH is mutex. Otherwise, it is the state (tH, s0, x0) where s0 = (s [ a2AH Dela) [ [ a2AH Adda (2) and x0 is the result of applying the composition of the functions {NPFa | a 2 AH} to x. Note that the happenings consist of grounded actions, i.e., all operator parameters are instantiated with constants. To introduce the semantics of derived predicates, we now modify the result of executing the happening. (We will also adapt the definition of mutex actions, see below.) The result of executing the happening is now obtained by applying the actions to s, then subtracting all derived facts from this, then applying the function D. That is, in the above definition we replace Equation 2 with the following: s0 = D(((s [ a2AH Dela) [ [ a2AH Adda) D) (3) 526 The Deterministic Part of IPC-4: An Overview where D denotes the set of all derived facts. If there are no derived predicates, D is the empty set and D is the identity function. As an example, say we have a Blocksworld instance where A is on B is on C, s ={ clear(A), on(A, B), on(B, C), ontable(C), above(A, B), above(B, C), above(A, C)}, and our happening applies an action that moves A to the table. Then the happening execution result will be computed by removing on(A, B) from s, adding clear(B) and ontable(A) into s, removing all of above(A, B), above(B, C), and above(A, C) from s, and applying D to this, which will re-introduce (only) above(B, C). So s0 will be s0 = {clear(A), ontable(A), clear(B), on(B, C), ontable(C), above(B, C) }. By the definition of happening execution, Fox and Long (2003) define the state transitions in a plan. The definitions of what an executable plan is, and when a plan achieves the goal, are then standard. The plan is executable if the result of all happenings in the plan is defined. This means that all action preconditions have to be fulfilled in the state of execution, and that no two pairs of actions in a happening are mutex. The plan achieves the goal if the goal holds true in the state that results after the execution of all actions in the plan. With our above extension of the definition of happening executions, the definitions of plan executability and goal achievement need not be changed. We do, however, need to adapt the definition of when a pair of actions is mutex. This is important if the happenings can contain more than one action, i.e., if we consider parallel (Graphplan-style) or concurrent (durational) planning. Fox and Long (2003) give a conservative definition that forbids the actions to interact in any possible way. The definition is the following. Definition 12 Mutex Actions (Fox & Long, 2003)Two grounded actions, a and b are non-interfering if GP rea " (Addb [ Delb) = GP reb " (Adda [ Dela) = ;, Adda " Delb = Addb " Dela = ;, La " Rb = Ra " Lb = ;, La " Lb ` L*a [ L*b (4) If two actions are not non-interfering they are mutex. Note that the definition talks about grounded actions where all operator parameters are instantiated with constants. La, Lb, Ra, and Rb refer to the left and right hand side of a's and b's numeric effects. Adda/Addb and Dela/Delb are a's and b's positive (add) respectively negative (delete) effects. GP rea/Gpreb denotes all (ground) facts that occur in a's/b's precondition. If a precondition contains quantifiers then these are grounded out (8x transforms to V ci, 9x transforms to W ci where the ci are all objects in the given instance), and GP re is defined over the resulting quantifier-free (and thus variable-free) formula. Note that this definition of mutex actions is very conservative - if, for example, fact F occurs only positively in a's precondition, then it does not matter if F is among the add effects of b. The conservative definition has the advantage that it makes it algorithmically very easy to figure out if or if not a and b are mutex. In the presence of derived predicates, the above definition needs to be extended to exclude possible interactions that can arise indirectly due to derived facts, in the precondition of the one action, whose truth value depends on the truth value of (basic) facts affected 527 Hoffmann & Edelkamp by the effects of the other action. In the same spirit in that Fox and Long (2003) forbid any possibility of direct interaction, we now forbid any possibility of indirect interaction. Assume we ground out all rules (P (x), OE(x)) for the derived predicates, i.e., we insert all possible vectors c of constants; we also ground out the quantifiers in the formulas OE(c), ending up with variable free rules. We define a directed graph where the nodes are (ground) facts, and an edge from fact F to fact F 0 is inserted iff there is a grounded rule (P (c), OE(c)) such that F 0 = P (c), and F occurs in OE(c). Now say we have an action a, where all ground facts occurring in a's precondition are, see above, denoted by GP rea. By DP rea we denote all ground facts that can possibly influence the truth values of the derived facts in GP rea: DP rea := {F | there is a path from F to an F 0 2 GP rea} (5) The definition of mutex actions is now updated simply by replacing Equation 4 with: (DP rea [ GP rea) " (Addb [ Delb) =( DP reb [ GP reb) " (Adda [ Dela) = ;, Adda " Delb = Addb " Dela = ;, La " Rb = Ra " Lb = ;, La " Lb ` L*a [ L*b (6) Note that the only thing that has changed is the first line, regarding interference of propositional effects and preconditions. As an example, reconsider the Blocksworld and the "above" predicate. Assume that the action that moves a block A to the table requires as an additional, derived, precondition, that A is above some third block. Then, in principle, two actions that move two different blocks A and B to the table can be executed in parallel. Which block A (B) is on can influence the above relations in that B (A) participates; however, this does not matter because if A and B can be both moved then this implies that they are both clear, which implies that they are on top of different stacks anyway. We observe that the latter is a statement about the domain semantics that either requires non-trivial reasoning, or access to the world state in which the actions are executed. In order to avoid the need to either do non-trivial reasoning about domain semantics, or resort to a forward search, our definition is the conservative one given above. The definition makes the actions moving A and B mutex on the grounds that they can possibly influence each other's derived preconditions. The definition adaptations described above suffice to define the semantics of derived predicates for the whole of PDDL2.2. Fox and Long (2003) reduce the temporal case to the case of simple plans above, so by adapting the simple-plan definitions we have automatically adapted the definitions of the more complex cases. In the temporal setting, PDDL2.2 level 3, the derived predicates semantics are that their values are computed anew at each happening in the plan where an action effect occurs. As said, in IPC-4 we used derived predicates only in the non-temporal setting. Some remarks on limitations of the IPC-4 treatment of derived predicates, and on future prospects, are in Section 7. 3.2 Timed Initial Literals Timed initial literals are a syntactically very simple way of expressing a certain restricted form of exogenous events: facts that will become TRUE or FALSE at time points that are known to the planner in advance, independently of the actions that the planner chooses 528 The Deterministic Part of IPC-4: An Overview to execute. Timed initial literals are thus deterministic unconditional exogenous events. Syntactically, we simply allow the initial state to specify - beside the usual facts that are true at time point 0 - literals that will become true at time points greater than 0. Timed initial literals are practically very relevant: in the real world, deterministic unconditional exogenous events are very common, typically in the form of time windows - within which a shop has opened, within which humans work, within which traffic is slow, within which there is daylight, within which a seminar room is occupied, within which nobody answers their mail because they are all at conferences, etc. The timed initial literals syntax is just about the simplest way one can think of to communicate such things to a planner. Timed initial literals can easily be compiled into artificial constructs (Fox, Long, & Halsey, 2004), involving an only linear blow-up in the instance representation and in plan length (i.e., number of actions in it). Still it seems highly likely that handing the timed literals over to an automated planner explicitly results in far better performance than when one hands over the artificial (compiled) representation. The results obtained in IPC-4 confirm this, see also Section 5. 3.2.1 Syntax The BNF notation is: hiniti ::= (:init hinit-eli*) hinit-eli ::=:timed-initial-literals (at hnumberi hliteral(name)i) The requirement flag for timed initial literals implies the requirement flag for durational actions, i.e., as said the language construct is only available in PDDL2.2 level 3. The timesh numberi at which the timed literals occur are restricted to be greater than 0. If there are also derived predicates in the domain, then the timed literals are restricted to not influence any of these, i.e., like action effects they are only allowed to affect the truth values of the basic (non-derived) predicates (IPC-4 will not use both derived predicates and timed initial literals within the same domain). As an illustrative example, consider a planning task where the goal is to have completed the shopping. There is a single action go-shopping that achieves the goal, and requires the (single) shop to be open as the precondition. The shop opens at time 9 relative to the initial state, and closes at time 20. We can express the shop opening times by two timed initial literals: (:init (at 9 (shop-open)) (at 20 (not (shop-open))) ) 3.2.2 Semantics We now describe the updates that need to be made to the PDDL2.1 semantics definitions given by Fox and Long (2003). Adapting two of the definitions suffices. 529 Hoffmann & Edelkamp The first definition we need to adapt is the one that defines what a "simple plan", and its happening sequence, is. The original definition by Fox and Long (2003) is this. Definition 11 Simple Plan (Fox & Long, 2003) A simple plan, SP , for a planning instance, I, consists of a finite collection of timed simple actions which are pairs (t, a), where t is a rational-valued time and a is an action name. The happening sequence, {ti}i=0...k for SP is the ordered sequence of times in the set of times appearing in the timed simple actions in SP . All ti must be greater than 0. It is possible for the sequence to be empty (an empty plan). The happening at time t, Et, where t is in the happening sequence of SP , is the set of (simple) action names that appear in timed simple actions associated with the time t in SP . In the STRIPS case, the time stamps are the natural numbers 1, . . . , n when there are n actions/parallel action sets in the plan. The happenings then are the actions/parallel action sets at the respective time steps. Fox and Long (2003) reduce the temporal planning case to the simple plan case defined here by splitting each durational action up into at least two simple actions - the start action, the end action, and possibly several actions in between that guard the durational action's invariants at the points where other action effects occur. So in the temporal case, the happening sequence is comprised of all time points at which "something happens", i.e., at which some action effect occurs. To introduce our intended semantics of timed initial literals, all we need to do to this definition is to introduce additional happenings into the temporal plan, namely the time points at which some timed initial literal occurs. The timed initial literals can be interpreted as simple actions that are forced into the respective happenings (rather than selected into them by the planner), whose precondition is true, and whose only effect is the respective literal. The rest of Fox and Long (2003)'s definitions then carry over directly (except goal achievement, which involves a little care, see below). The PDDL2.2 definition of simple plans is this here. Definition 11 Simple Plan A simple plan, SP , for a planning instance, I, consists of a finite collection of timed simple actions which are pairs (t, a), where t is a rational-valued time and a is an action name. By tend we denote the largest time t in SP , or 0 if SP is empty. Let T L be the (finite) set of all timed initial literals, given as pairs (t, l) where t is the rational-valued time of occurrence of the literal l. We identify each timed initial literal (t, l) in T L with a uniquely named simple action that is associated with time t, whose precondition is TRUE, and whose only effect is l. The happening sequence, {ti}i=0...k for SP is the ordered sequence of times in the set of times appearing in the timed simple actions in SP and T L. All ti must be greater than 0. It is possible for the sequence to be empty (an empty plan). The happening at time t, Et, where t is in the happening sequence of SP , is the set of (simple) action names that appear in timed simple actions associated with the time t in SP or T L. Thus the happenings in a temporal plan are all points in time where either an action effect, or a timed literal, occurs. The timed literals are simple actions forced into the plan. With this construction, Fox and Long (2003)'s Definitions 12 (Mutex Actions) and 13 530 The Deterministic Part of IPC-4: An Overview (Happening Execution), as described (and adapted to derived predicates) in Section 3.1.2, can be kept unchanged. They state that no action effect is allowed to interfere with a timed initial literal, and that the timed initial literals are true in the state that results from the execution of the happening they are contained in. Fox and Long (2003)'s Definition 14 (Executability of a plan) can also be kept unchanged - the timed initial literals change the happenings in the plan, but not the conditions under which a happening can be executed. The only definition we need to reformulate is that of what the makespan of a valid plan is. In Fox and Long (2003)'s original definition, this is implicit in the definition of valid plans. The definition is this. Definition 15 Validity of a Simple Plan (Fox & Long, 2003) A simple plan (for a planning instance, I) is valid if it is executable and produces a final state S, such that the goal specification for I is satisfied in S. The makespan of the valid plan is accessible in PDDL2.1 and PDDL2.2 by the "totaltime" variable that can be used in the optimization expression. Naturally, Fox and Long (2003) take the makespan to be the end of the plan, the time point of the plan's final state. In the presence of timed initial literals, the question of what the plan's makespan is becomes a little more subtle. With Fox and Long (2003)'s above original definition, the makespan would be the end of all happenings in the simple plan, which include all timed initial literals (see the revised Definition 11 above). So the plan would at least take as long as it takes until no more timed literals occur. But a plan might be finished long before that - imagine something that needs to be done while there is daylight; certainly the plan does not need to wait until sunset. We therefore define the makespan to be the earliest point in time at which the goal condition becomes (and remains) true. Formally this reads as follows. Definition 15 Validity and Makespan of a Simple Plan A simple plan (for a planning instance, I) is valid if it is executable and produces a final state S, such that the goal specification for I is satisfied in S. The plan's makespan is the smallest t >= tend such that, for all happenings at times t0 >= t in the plan's happening sequence, the goal specification is satisfied after execution of the happening. Remember that tend denotes the time of the last happening in the plan that contains an effect caused by the plan's actions - in simpler terms, tend is the end point of the plan. What the definition says is that the plan is valid if, at some time point t after the plan's end, the goal condition is achieved and remains true until after the last timed literal has occurred. The plan's makespan is the first such time point t. Note that the planner can "use" the events to achieve the goal, by doing nothing until a timed literal occurs that makes the goal condition true - but then the waiting time until the nearest such timed literal is counted into the plan's makespan. (The latter is done to avoid situations where the planner could prefer to wait millions of years rather than just applying a single action itself.) Remember that the makespan of the plan, defined as above, is what can be denoted by total-time in the optimization expression defined with the problem instance. As with the derived predicates, the definition adaptations above suffice to define the semantics of timed initial literals for PDDL2.2. Some remarks on limitations of the language construct, and on future prospects, are in Section 7. 531 Hoffmann & Edelkamp 4. The Competitors We now provide an overview table listing the systems along with their language capabilities (subset of PDDL2.2 covered), and we sketch the competing systems. The deterministic part of IPC-4 attracted 19 competitors (21 when counting different system versions), 12 (14) of which were satisficing and 7 of which were optimal. Each competitor wrote a 2- 3 page extended abstract for inclusion in the IPC-4 booklet at ICAPS 2004 (Edelkamp et al., 2004). The included sketches are very brief outlines of these abstracts. The reader is encouraged to read the original work by the system authors. At this point it is appropriate to say a few words on system development. We allowed competitors to modify their systems during the competition phase, i.e., while data collection was running. Our reasons for doing so were twofold. First, some of our domains were quite unusual - most particularly, those containing ADL constructs compiled to STRIPS - so we (rightly) expected planner implementations to have parsing etc. trouble with them. Second, from our own experience with the competitions we knew that people will try to enhance their planners anyway so we did not see much point in forbidding this. We trusted the competitors to not do stupid things like hard-coding domain names etc., and we trusted to have a diverse enough range of domains to make tuning a "domain-independent" task. An alternative would have been to collect executables before data collection, and then have all the results collected by the organizers. Anyone who has ever run experiments with someone else's planner knows that such an approach is completely infeasible due to the prototype-nature of the systems, and due to the many tiny details in minor language changes, programming environments, etc. - if one wants to obtain meaningful results, at least. One could in principle apply a strict "any failure is counted as unsolved" rule, but doing so would likely lay way too much emphasis on little programming errors that have nothing to do with the evaluated algorithms. The competition phase was, roughly, from February 2004 until middle May 2004. During that time, we released the domains one-by-one, and the competitors worked on their systems, handing the results over to us when they were ready. At the end of that phase, the competitors had to submit an executable, and we ran some sampled tests to see that the executable did really produce the reported results, across all the domains. The abstracts describing the systems had to be delivered a little earlier, by the end of April, due to timing constraints for printing the booklet. We start with the overview table, then sketch the satisficing and optimal competitors. 4.1 Planner Overview An overview of the participating satisficing planners, and their language capabilities (defined as what language features they attacked in IPC-4), is given in Table 1. Observe that most of the planners treat only a small subset of PDDL2.2: from a quick glance at the table, one sees that there are far more ("-") entries than ("+") entries. Note here that, often, even when a ("+") sign is given, a planner may treat only a subset of the specified language feature (as needed for the respective IPC-4 domain). For example, a planner might treat only a subset of ADL, or only linear arithmetic for numeric variables. We leave out these details for the sake of readability. 532 The Deterministic Part of IPC-4: An Overview ADL DP Numbers Durations TL CRIKEY - - + + - FAP - - - - - FD, FDD + + - - - LPG-TD + + + + + Macro-FF + - - - - Marvin + + - - - Optop + + + + + P-MEP + - + + + Roadmapper - - - - - SGPlan + + + + + Tilsapa - - + + + YAHSP - - - - - BFHSP - - - - - CPT - - - + - HSP*a - - + + - Optiplan - - - - - SATPLAN - - - - - SemSyn + - + - - TP4 - - + + - Table 1: An overview of the planners in IPC-4, and their language capabilities (i.e., the language features they attacked in IPC-4). satisficing planners are in the top half of the table, optimal planners are in the bottom half. Each table entry specifies whether ("+") or not ("-") a planner can handle a language feature. "DP" is short for derived predicates, "TL" is short for timed initial literals. With "Numbers" and "Durations" we mean numeric fluents and fixed action durations (no duration inequalities), in the sense of PDDL2.1 level 2 and 3, respectively. The planners (and their name abbreviations) are explained below. LPG-TD, Optop, and SGPlan are the only planners that treat the full range of PDDL2.2, as used in IPC-4. Three satisficing planners (FAP, Roadmapper, and YAHSP), and three optimal planners (BFHSP, Optiplan, and SATPLAN), treat only pure STRIPS. Derived predicates are treated by 4 planners, timed initial literals by 5 planners, ADL by 7 planners, numeric variables by 8 planners, and action durations by 9 planners. Table 1 shows that there is a significant amount of acceptance of the new (PDDL2.1 and PDDL2.2) language features, in terms of implemented systems participating in the IPC. Table 1 also shows that the development on the systems side has a tendency to be slower than the development on the IPC language side: even the most wide-spread language feature beyond STRIPS, action durations, is dealt with by less than half of the 19 planners. In our opinion, this should be taken to indicate that further language extensions should be made slowly. Some more on this is said in Section 7. 533 Hoffmann & Edelkamp 4.2 Satisficing Planners The satisficing planners - the planners giving no guarantee on the quality of the returned plan - were the following. We proceed in alphabetical order. CRIKEY by Keith Halsey, University of Strathclyde, Glasgow, UK. CRIKEY is a heuristic search forward state space planner. It includes a heuristic for relaxed plans for temporal and numeric problems and applies a scheduler based on simple temporal networks to allow posterior plan scheduling. FAP by Guy Camilleri and Joseph Zalaket, University Paul Sabatier / IRIT CCICSC, France. FAP handles non-temporal non-numeric domains. It is a heuristic planner using the relaxed planning heuristic in an N -best forward search, with meta-actions (action sequences) extracted from the relaxed planning graph to perform jumps in the state space. Fast Downward, FD for short, and Fast Diagonally Downward, FDD for short, by Malte Helmert and Silvia Richter, University of Freiburg, Germany. FD and FDD can treat non-temporal non-numeric domains. They apply a new heuristic estimate based on a polynomial relaxation to the automatically inferred multivariate or SAS+ representation of the problem space. FDD also applies the traditional FF-style relaxed planning heuristic, i.e., it applies both heuristics in a hybrid search algorithm. LPG-TD by Alfonso Gerevini, Alessandro Saetti, Ivan Serina, and Paolo Toninelli, University of Brescia, Italy. The extensions to the randomized local plan graph search that was already included in LPG at IPC-3 includes functionality for PDDL2.2 derived predicates and timed initial literals, as well as various implementation refinements. A version tailored to computation speed, LPG-TD.speed, and a version tailored for plan quality, LPG-TD.quality, participated. LPG-TD.quality differs from LPG-TD.speed basically in that it does not stop when the first plan is found but continues until a stopping criterion is met. The LPG-TD team also ran a third version, called "LPG-TD.bestquality", which used, for every instance, the entire half hour CPU time available to produce as good a plan as possible. This third version is not included in the official data because every team was allowed to enter at most two system versions. Macro-FF by Adi Botea, Markus Enzenberger, Martin M"uller, Jonathan Schaeffer, University of Alberta, Canada. As the name suggests Macro-FF extends Hoffmann's FF planner with macro operators (and other implementation refinements). Macros are learned prior to the search and fed into the planner as new data, separated from the operator file and the problem file. At runtime, both regular actions and macro-actions are used for state expansion. Heuristic rules for pruning instantiations of macros are added to FF's original strategy for search control. Marvin by Andrew Coles and Amanda Smith, University of Strathclyde, Glasgow, UK. Marvin extends FF by adding extra features and preprocessing information, such as plateau-escaping macro actions, to enhance the search algorithm. Optop by Drew McDermott, Yale University, USA. Optop is an extension of the wellknown UNPOP planner, with the ability to handle a complex input language including, amongst other things, autonomous processes running in parallel to the actions taken by the planner. The underlying principle is a forward search with heuristic guidance obtained from greedy-regression match graphs that are built backwards from the goals. 534 The Deterministic Part of IPC-4: An Overview P-MEP by Javier Sanchez, Minh Tang, and Amol D. Mali, University of Wisconsin, USA. P-MEP is short for Parallel More Expressive Planner. Unlike most planners, P-MEP can treat numeric variables with non-linear action effects. It employs a forward search with relaxed plan heuristics, enhanced by relevance detection in a pre-process, and by taking into account exclusion relations during relaxed planning. Roadmapper by Lin Zhu and Robert Givan, Purdue University, USA. Roadmapper handles non-temporal non-numeric domains. It is a forward heuristic search planner enhancing the FF heuristic with a reasoning about landmarks. The latter are propositions that must be true at some point in every legal plan. Roadmapper finds landmarks in a pre-process, arranges them in a directed road-map graph, and uses that graph to assign weights to actions in the FF-style relaxed plans. SGPlan by Yixin Chen, Chih-Wei Hsu and Benjamin W. Wah, University of Illinois, USA. The planner bootstraps heuristic search planners by applying Lagrange optimization to combine the solution of planning subproblems. To split the problem, an ordering of the planning goals is derived. The incremental local search strategy that is applied for Lagrange optimization on top of the individual planners relies on the theory of extended saddle points for mixed integer linear programming. Tilsapa by Bharat Ranjan Kavuluri and Senthil U., AIDB Lab, IIT Madras, India. The planner extends the SAPA system that handles temporal numeric domains, with the ability to handle timed initial literals. YAHSP by Vincent Vidal, University of Artois, France. YAHSP, an acronym for yet another heuristic search planner, searches forward with a FF-style relaxed planning heuristic. The main enhancement is a relaxed-plan-analysis phase that replaces actions in the relaxed plan based on heuristic notions of better suitability in reality, producing macro-actions in the process. The macro-actions are, basically, as long as possible feasible sub-sequences of the (modified) relaxed plan. 4.3 Optimal Planners The participating optimal planners were the following, in alphabetical order. BFHSP by Rong Zhou and Eric A. Hansen, Mississippi State University, USA. BFHSP, for breadth-first heuristic search planner, optimizes the number of actions in the plan. The planner implements the standard max-atom and max-pair heuristics (as well as the max-triple heuristic, which was however not used in IPC-4). The main difference to other systems is its search algorithm called breadth-first heuristic search; this improves the memory requirements of A* search by searching the set of nodes up to a certain threshold value in breadth-first instead of best-first manner. CPT by Vincent Vidal, University of Artois, France, and Hector Geffner, University of Pompeu Fabra, Barcelona, Spain. CPT optimizes makespan. The planner is based on constraint satisfaction, and transforms the planning problem to a CSP. The branching scheme that solves the CSP makes use of several constraint propagation techniques related to POCL, and of the temporal max-atom heuristic. HSP*a by Patrik Haslum, Link"oping University, Sweden. HSP*a is a derivate of TP4, see below, with the same expressivity, but with a weakened version of the MaxTriple heuristic instead of the MaxPair heuristic. 535 Hoffmann & Edelkamp Optiplan by Menkes van der Briel and Subbarao Kambhampati, Arizona State University, USA. Optiplan is a planner based on integer programming (IP), which itself is a consequent extension to the encoding of the planning graph used in the planning as satisfiability approach (see below). The compiled planning problem is solved using the CPLEX/ILOG system. An interesting property of Optiplan is that, due to the power of the underlying IP solver, it can optimize a great variety of different plan metrics, in fact every plan metric that can be expressed with a linear function. In that respect, the planner is unique. In IPC-4, step-optimal plans were computed because that is typically most efficient. SATPLAN04, SATPLAN for short, by Henry Kautz, David Roznyai, Farhad TeydayeSaheli, Shane Neph, and Michael Lindmark, University of Washington, USA. SATPLAN treats non-temporal non-numeric domains and optimizes the number of parallel time steps. The system did not participate in IPC-3 but in IPC-1 and IPC-2, so it was a real comeback. As the planner compiles a planning problem to a series of satisfiability tests, the performance of the system relies on the integrated back-end SAT solver. The graph-plan encoding scheme underwent only minor changes, but the underlying SAT solver is much more powerful than that of 4 years ago. SemSyn by Eric Parker, Semsyn Software, Tempe, Arizona. The SemSyn system optimizes the number of actions. Semsyn is linked to a commercial product and applies a combination of forward and backward chaining. Both searches are variants of the original algorithms. Forward chaining is goal-directed, while backward chaining is generalized, comprising a partition of the top-level goal. As the product is proprietary, detailed information on the system is limited. TP4-04, TP4 for short, by Patrik Haslum, Link"oping University, Sweden. The planner is an extended makespan-optimal scheduling system that does an IDA* search routine with the max-pair heuristic. It can deal with a restricted form of metric planning problems with numerical preconditions and effects. 5. Results The CPU times for IPC-4 were measured on a machine located in Freiburg, running 2 Pentium-4 CPUs at 3 GHz, with 6 GB main memory. Competitors logged in remotely and ran their planners, producing one separate ASCII result file for each solved instance, giving the runtime taken, the quality of the found plan, as well as the plan itself. Each planner run was allowed at most half an hour (CPU, not real) runtime, and 1 GB memory, i.e., these were the cutoff values for solving a single instance. As a kind of performance measure, the IPC-3 version of LPG (Gerevini, Saetti, & Serina, 2003) (called "LPG-3"), i.e. the most successful automatic planner from IPC-3, was also run (by its developers, on top of the workload with their new planner version LPG-TD). We remark that, for the sake of simplicity, to save CPU time, and to ensure fairness, randomized planners were allowed only a single run one ach instance, with a fixed random seed. IPC-4 featured a lot of domain versions and instances, and a lot of planners. The only way to fully understand the results of such a complex event is to examine the results in detail, making sense of them in combination with the descriptions/PDDL encodings of the domains, and the techniques used in the respective planners. We recommend doing so, at least to some extent, to everybody who is interested in the results of (the deterministic part 536 The Deterministic Part of IPC-4: An Overview of) IPC-4. The on-line appendix of this paper includes all the individual solution files. The appendix also includes all gnuplot graphics we generated (more details below), as well as a GANNT chart for each of the result files; the GANNT charts can be visualized using the Vega visualization front-end (Hipke, 2000). In what follows, we provide an overview of the results, and we highlight (what we think are) the most important points; we explain how we decided about the IPC-4 competition awards. More precisely, Section 5.1 gives an overview over the results, in terms of percentages of attacked and solved instances per language subset and planner. Section 5.2 describes how we evaluated the results, particularly what procedure we chose to decide about the awards. Thereafter, Sections 5.3 to 5.9 in turn consider in some more detail the results in the single domains. 5.1 Results Overview One crude way to obtain an overview over such a large data set is to simply count the number of instances each planner attacked, i.e. tried to solve, and those that it succeeded to solve. In our case, it also makes sense to distinguish between the different PDDL2.2 subsets used by the IPC-4 benchmarks. The data is displayed in Table 2. Table 2 is complicated, and needs some explanation. First, consider the columns in the table. The leftmost column is, as usual, for table indexing. The rightmost column contains data for the entire set of instances in IPC-4. The columns in between all refer to a specific subset of the instances, in such a way that these subsets are disjunct and exhaustive - i.e., the instance sets associated with columns are pairwise disjunct, and their union is the entire set of instances. The subsets of instances are defined by the PDDL2.2 subsets they use. The abbreviations in Table 2 are explained in the caption. "X+Y" means that these instances use language features "X" and "Y", and none of the other features. For example, "N+DP" are (only) those instances with uniform durations (i.e., PDDL2.2 level 1) and derived predicates. Note that Table 2 does not have a column for all possible combinations of language features - we show only those that were used by (a non-empty subset of) the IPC-4 benchmarks. Note also that we do not distinguish between different domain formulations, i.e. between ADL and STRIPS. The instance numbers in Table 2 are counted for domain versions. That is, if there is an ADL and a STRIPS formulation of a version, then each instance is counted only once.4 The first line in Table 2, indexed "Number", simply gives the size of the instance set associated with the column. All other lines correspond, obviously, to planning systems. The upper half of the table contains, in alphabetical order, the satisficing planners; the lower half contains the optimal planners. Note that we list only one version of LPG-TD. This is because the "speed" and "quality" version of this planner differ in terms of plan quality, but not in terms of the number of attacked/solved instances. 4. As said, each domain featured several domain versions, differing in terms of the application constraints modelled. Within each domain version, there could be several different domain version formulations, differing in terms of the precise PDDL subset used to encode the same semantics. In most cases with more than one formulation, there were exactly two formulations, using/not using ADL constructs, respectively. Competitors could choose the formulation they liked best and the results across different formulations were evaluated together. 537 Hoffmann & Edelkamp N N+DP N+NV D D+TL D+NV D+TL+NV All Number 382 196 152 302 116 272 136 1,556 CRIKEY 42 77 -- 47 66 -- 98 55 -- 56 61 42 FAP 33 64 -- -- -- -- -- -- 33 74 16 FD 83 62 84 100 -- -- -- -- -- 83 75 28 FDD 92 62 84 100 -- -- -- -- -- 88 75 28 LPG-3 55 62 -- 42 24 45 62 -- 56 50 -- 51 54 38 LPG-TD 67 87 75 74 61 37 76 62 63 100 96 50 87 100 76 71 71 Macro-FF 57 87 -- -- -- -- -- -- 57 100 21 Marvin 67 62 34 100 -- -- -- -- -- 52 75 28 Optop -- -- 8 43 -- -- 8 7 3 P-MEP 15 61 -- 8 55 24 45 24 43 13 32 17 43 38 Roadmapper 28 49 -- -- -- -- -- -- 28 56 12 SGPlan 68 100 65 100 64 100 75 90 78 74 85 100 74 100 73 96 96 Tilsapa -- 10 69 62 63 38 13 11 YAHSP 77 87 -- -- -- -- -- -- 77 100 21 BFHSP 29 87 -- -- -- -- -- -- 29 100 21 CPT 33 87 -- -- 22 100 -- -- -- 28 100 41 HSP*a 33 38 -- 10 62 -- 50 50 -- 29 44 30 Optiplan 34 87 -- -- -- -- -- -- 34 100 21 SATPLAN 46 87 -- -- -- -- -- -- 46 100 21 SemSyn 49 48 -- 11 37 -- -- -- -- 40 23 15 TP4 29 64 -- 17 62 -- 52 50 -- 31 54 37 Table 2: An overview of the IPC-4 results. Used abbreviations: "N" no durations, "DP" derived predicates, "NV" numeric variables, "D" durations, "TL" timed initial literals. Line "Number" gives number of instances, all other lines give percentage values. In each entry, the right number is the percentage of instances that were attacked, and the left number is the percentage of these that were solved - i.e. the left number is the success ratio. In the rightmost column, the middle number is the percentage of attacked instances relative to all instances that lie within the language range of the respective planner; the right number is the percentage of attacked instances over all. A dash indicates that a planner can not handle a language subset. See the detailed explanation in the text. Let us consider the table entries in between the leftmost and rightmost columns, i.e., the entries concerning a planner "X" and a language/instance subset "Y". The numbers in these entries are the success ratio and the attacked ratio. Precisely, they were obtained as follows. We first counted the number "y" of instances in "Y" that planner "X" tried to solve - our definition was to take all domain versions (inside "Y") for which "X" delivered results, and set "y" to be the total number of instances in these domain versions. We then counted the number "x" of instances in "Y" that planner "X" succeeded to solve. We obtained the first number in the table entry - the success ratio - as the ratio (in percent) of "x" divided by "y". We obtained the second number in the entry - the attacked ratio - as the ratio (in percent) of "y" divided by the size of "Y". For space reasons, we rounded the percentages 538 The Deterministic Part of IPC-4: An Overview to the values shown in the table. A dash in the table entry means that the planner "X" can not handle the language subset "Y". An empty table entry means that the planner can handle "Y", but did not attack any instance in "Y". In a few cases, namely in the Promela domain, see the discussion in Section 5.5, we could not formulate the largest instances in STRIPS because these (fully-grounded) representations became too large. So, there, the numbers of test instances are different between the ADL and STRIPS formulations of the same domain version.5 Table 2 uses the number of ADL instances, not taking account of this subtlety. This implies a slight disadvantage in terms of success ratio in columns "N" and "All", for the planners that attacked the smaller STRIPS test suites. These planners are the following, with correct column "N" success ratio in parentheses: CRIKEY (51), FAP (41), LPG-TD (74), SGPlan (80), YAHSP (92), BFHSP (35), CPT (39), HSP*a (52), Optiplan (41), SATPLAN (55), and TP4 (37). The table entries in the rightmost column are obtained similarly as above. They summarize the situation regarding the respective planners across all used language subsets. The left and right number in each entry give the ratio of the number of solved instances to the number of attacked instances, and the number of attacked instances to the number of all instances, respectively. The number in the middle of each entry gives the ratio of the number of attacked instances to the number of all instances that lie within the language range of the respective planner. We included this number in order to provide a measure of to what extent each planner/team attacked those instances that they could attack. Some remarks are in order regarding the latter ratio, and the ratio of attacked instances in general. First, there was no rule telling the competitors to "attack everything they can" - competitors were free to choose. In retrospect, we feel that there should have been such a rule, since it would make data interpretation easier. The way things are, it is not known for what reason a planner did not attack an instance subset (domain version): Bad results? Not interested? Overlooked? Second, many planners can handle only subsets of certain language features, like, of numeric variables ("NV"). This can lead to too low percentages in Table 2, where for simplicity we do not take account of these details. Third, one detail we did take account of is that 50 of the 382 instances in column "N" (namely, the "middle-compiled" version of the PSR domain, see Section 5.6) are formulated in ADL only, and are thus not accessible to many of the planners. For planners not able to handle ADL, when computing the middle number in the rightmost table entry, we subtracted these 50 instances from the "instances within the language range" of the respective planner. In column "N", however, we took the usual ratio to the set of all instances - in particular, this is why Macro-FF, YAHSP, BFHSP, CPT, Optiplan, and SATPLAN all have an 87% attacked-ratio in the "N" column, but a 100% attacked-ratio as the middle number of the rightmost column. All in all, the planners all have a pretty good coverage of the language subset they could attack; except Optop, Tilsapa, and maybe Semsyn. These planners/teams were probably interested only in a small subset of the instances.6 In the other cases where instances 5. In PSR, it also was impossible to formulate the largest instances in STRIPS. But, there, we split the domain into different versions regarding the size and formulation language of the instances. We did not want to introduce more distinction lines in Promela because there we already had 8 different domain versions. 6. Drew McDermott, the developer of Optop, told us in private conversation that he tried to solve only the most complicated domain versions. 539 Hoffmann & Edelkamp were left out, planners/teams typically did not attack the domain versions where derived predicates or timed initial literals had been compiled away. Note that a lack of coverage may also be due to language details not accounted for by the rather crude distinctions made for Table 2.7 What can we conclude from Table 2, in terms of planner performance? Some spotlights are these: * The success ratios of optimal planners are generally lower than those of satisficing planners. The largest overall success ratio of an optimal planner, SATPLAN, is 46 (55 when taking account of the above mentioned subtlety regarding ADL/STRIPS formulations in Promela); compared to 88 for FDD. * However, there are also various cases where an optimal planner has a higher success ratio than a satisficing one. * FD, FDD, and YAHSP have the best success ratios (as mentioned, the success ratio of YAHSP in the "N" column is 92 when taking account of ADL/STRIPS in Promela). FD and FDD also have a pretty good coverage in the "N" and "N+DP" columns, i.e. within PDDL2.2 level 1 (the left out instances are the domain versions with compiled derived predicates). * SGPlan has an extremely high coverage, attacking 96% of all instances. Even so, it has a very competitive success ratio. (Concretely, SGPlan solved 1,090 of the 1,556 IPC-4 instances; it attacked 1,496. The second place in terms of the number of instances solved is held by LPG-TD, which solved 843 out of the 1108 instances it could attack.) * LPG-3, i.e. the IPC-3 version of LPG, ranges somewhere in the middle-ground of the planners, in terms of success ratio and coverage. This indicates that significant performance improvements were made in difference to IPC-3. (Note that this holds for the new LPG version just as for the other top performance planners.) Naturally, in our evaluation of the results, particularly in the evaluation that formed the basis of the award decisions, we undertook a more detailed examination of the data set. We considered, as much as possible, scaling behavior, rather than the simple instance counts above. In what follows, we first explain our evaluation process in detail, then we consider each of the domains in turn. 5.2 Results Evaluation As said, runtime data for satisficing and optimal planners was considered separately. For plan quality, we distinguished between three kinds of optimization criteria: number of actions, makespan (equalling the number of time steps in the non-temporal domains), and metric value. The last of these three was only defined if there was a :metric specification in the respective task. We compared each planner according to only a single plan quality criterion. That is, for every domain version the competitors told us what criterion their 7. In private conversation, the LPG-TD team told us that this is the case for LPG-TD, and that they in fact attacked all domain versions they could at the time IPC-4 was running. 540 The Deterministic Part of IPC-4: An Overview planner was trying to optimize in that domain version. We evaluated all (and only) those planners together that tried to optimize the same criterion. Our reason for doing so is that it does not make sense to evaluate planners based on criteria they don't actually consider. Altogether, for every domain version we created up to 5 gnuplot graphics: for satisficing runtime, for optimal runtime, for number of actions, for makespan, and for metric value. The plan quality data of optimal planners was put into the same plots as those for satisficing planners, to enable some comparison. In a few cases, we had to split a runtime figure (typically, for satisficing planners) into two separate figures because with too many runtime curves in it the graphic became unreadable. Based on the gnuplot graphics, we evaluated the data in terms of asymptotic runtime and solution quality performance. The comparisons between the planners were made by hand, i.e. by looking at the graphs. While this is simplistic, we believe that it is an adequate way of evaluating the data, given the goals of the field, and the demands of the event. It should be agreed that what we are interested in is the (typical) scaling behavior of planners in the specific domains. This excludes "more formal" primitive comparative performance measures such as counts of instances solved more efficiently - a planner may scale worse than another planner, yet be faster in a lot of smaller instances just due to pre-processing implementation details. The ideal formal measure of performance would be to approximate the actual scaling functions underlying the planners' data points. But it is infeasible to generate enough data, in an event like the IPC, to do such formal approximations. So we tried to judge the scaling behaviors by hand, laying most emphasis on how efficiently/if at all the largest instances in a test suite were solved by a planner.8 The rest of the section contains one sub-section for each domain in turn. Each subsection provides the following material. First, we give a brief description of the domain and its most important features (such as domain versions and formulations used in IPC-4). We then include a set of gnuplot graphics containing (what we think are) the most important observations regarding runtime (the total set of gnuplot graphs is too space-consuming). We discuss plan quality, with data comparing the relative performance of pairs of planners. We also add some intuitions regarding structural properties of the domains, and their possible influence on the performance of the planners. Finally, we provide the information - 1st and 2nd places, see below - underlying the decisions about the awards. This information is contained in a text paragraph at the end of the sub-section for each domain. As the basis for deciding about the awards, within every domain version, we identified a group of planners that scaled best and roughly similar. These planners were counted as having a 1st place in that domain version. Similarly, we also identified groups of 2nd place planners. The awarding of prizes was then simply based on the number of 1st and 2nd places of a planner, see Section 6.9 We consider the domains in alphabetical order. Before we start, there are a few more remarks to be made. First, some planners are left out of the plots in order to make the 8. Note that the success ratios displayed in Table 2 are also a crude way to access scaling behavior. 9. Of course, many of the decisions about 1st and 2nd places were very close; there were numerous special cases due to, e.g., what planners participated in what domain versions; summing up the places introduces a dependency of the results on the number of domain versions in the individual domains. To hand out awards one has to make decisions, and in these decisions a lot of detail is bound to disappear. One cannot summarize the results of such a huge event with just a few prizes. 541 Hoffmann & Edelkamp plots readable. Specifically, we left out LPG-TD.quality as well as LPG-3.quality (the IPC-3 version of LPG with a preference on quality); these planners are always slower than their counterparts with a preference on speed. We also left out FD, since in most cases it showed very similar behavior to FDD. We chose FDD because in a few cases its performance is superior. Note further that we do not distinguish the runtime performance of optimal planners optimizing different criteria. Finally, it can make a significant difference whether a planner is run on an ADL encoding, or on its compilation to STRIPS; this distinction also gets lost in our evaluation. We emphasize that we applied the simplifications only in order to improve readability and understandability of the results. We do not wish to imply that the planners/distinctions that are omitted aren't important. 5.3 Airport In the Airport domain, the planner has to control the ground traffic on an airport. The task is to find a plan that solves a specific traffic situation, specifying inbound and outbound planes along with their current and goal positions on the airport. The planes must not endanger each other, i.e. they must not both occupy the same airport "segment" (a smallest road unit), and if plane x drives behind plane y then between x and y there must be a safety distance (depending on the size of y). These safety constraints are modelled in terms of "blocked" and "occupied" predicates, whose value is updated and controlled via complex ADL preconditions and conditional effects. There were four different versions of the domain in IPC-4: a non-temporal version, a temporal version, a temporal version with time windows, and a temporal version with compiled time windows. The time windows in the latter two versions concern airplanes that are known to land in the future, and that will thus block certain airport segments (runways) during certain time windows. The time windows are modelled using timed initial literals, respectively their compilation. In every domain version, there is an ADL formulation, as well as a STRIPS formulation obtained by compiling the ADL constructs away (resulting in a partially grounded encoding). Instances scale in terms of the size of the underlying airport, as well as the number of airplanes that must be moved. One remarkable thing about the Airport domain is that the instances were generated based on a professional airport simulation tool. The largest instances used in IPC-4 (number 36 to 50) correspond to a real airport, namely Munich airport (MUC). Instance number 50 encodes a traffic situation with 15 moving airplanes, which is typical of the situations encountered at MUC in reality. Instances number 1 to 20 come from smaller toy airports, instances number 21 to 35 are based on one half of MUC airport. Figure 1 shows the runtime performance in the non-temporal version of the domain. For readability, the set of satisficing planners is split over two graphs. As will be the case in all graphs displayed in the subsequent discussions, the x-axis denotes the instance number (obviously, the higher the number the larger the instance), and the y-axis gives CPU runtime in seconds on a logarithmic scale. CPU time is total, including parsing and any form of static pre-processing. It can be observed in Figure 1 that FDD is the only planner solving all problem instances. LPG-TD and SGPlan both scale relatively well, but fail on the largest instances. 542 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDMACRO-FF SGPLANCRIKEY FDDYAHSP 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 FAPMARVIN PMEPROADMAPPER LPG-3 (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 BFHSPTP4 CPTOPTIPLAN SATPLANSEMSYN (c) Figure 1: Performance in non-temporal Airport, satisficing (a) and (b), optimal (c). The other planners all behave much more unreliably. Observe also that the IPC-3 version of LPG lags far behind FDD, LPG-TD, and SGPlan. For the optimal planners, which unsurprisingly behave clearly worse than the satisficing planners, the only clear-cut observation is a performance advantage for SATPLAN. The other optimal planners all behave relatively similarly; Semsyn and BFHSP are the only ones out of that group solving a few of the larger instances. For plan quality, there are two groups of planners, one trying to minimize the number of actions, and one trying to minimize makespan, i.e. the number of parallel action steps. In the former group, the plan quality performance differences are moderate. CRIKEY, LPGTD, SGPlan, and YAHSP sometimes find sub-optimal plans. FDD's plans are optimal in all cases where an optimal planner found a solution. As a measure of comparative plan quality, from now on we provide, given planners A and B, the min, mean, and maximum of the ratio quality(A) divided by quality(B), for all instances that are solved by both A and B. We call this data the "ratio A vs B". For CRIKEY (A) vs FDD (B), the ratio is 0.91 (min), 1.04 (mean), and 1.45 (max), [0.91(1.04)1.45] for short. For LPG-TD.speed vs FDD 543 Hoffmann & Edelkamp the ratio is [0.91(1.08)1.80]; for LPG-TD.quality vs FDD it is [0.91(1.06)1.80]. For SGPlan vs FDD it is [0.91(1.07)2.08]; for YAHSP vs FDD it is [0.91(1.07)1.43]. In the group of planners minimizing makespan, (only) Marvin has a tendency to find very long plans. The comparison Marvin vs the optimal SATPLAN is [1.00(2.46)4.64]. In the maximum case, SATPLAN finds a plan with 53 steps, and Marvin's plan is 246 steps long. An interesting observation concerns the two scaling parameters of that domain: the number of airplanes, and the size of the airport. In all the plots in Figure 1, and also in most plots in Figure 2 below, one can observe that the instances number 26 to 35 become increasingly hard for the planners - but in the step to instance number 36, the performance suddenly becomes better again. As said above, the instances 21 to 35 are based on one half of MUC airport, while the instances 36 to 50 are based on the full MUC airport. But between instances 26 and 35, the number of airplanes rises from 6 to 12; instance 36 contains only 2 airplanes, instances 37 and 38 contain 3. That is, it is easier for the planners to address a larger airport with fewer planes. Note here the domain combinatorics: the number of reachable states is in the order of Sn, where S is the number of different airport segments, and n is the number of airplanes. Figure 2 shows the runtime performance in the temporal versions of the domain. We first consider the domain versions without time windows: parts (a) and (b) of the figure. Of the satisficing planners, LPG-TD and SGPlan both scale just as well as they do in the non-temporal case; all other satisficing planners scale orders of magnitude worse. Of the optimal planners, CPT scales best, followed by TP4. The two planners minimizing the number of actions are SGPlan and CRIKEY; SGPlan behaves somewhat worse, the ratio is [0.95(1.03)1.54]. Of the planners that minimize makespan, only LPG-TD scales beyond the smallest instances. The ratio LPG-TD.speed vs the optimal CPT is [1.00(1.15)1.70], P-MEP vs CPT is [1.00(1.11)1.42]. The ratio LPG-TD.quality vs CPT is [1.00(1.03)1.44]. No optimal planner in the competition could handle timed initial literals, so with explicit time windows no such planner participated. With compiled time windows, only CPT participated, scaling a little worse than in the temporal domain version without time windows. For the satisficing planners, the results are more interesting. With explicit time windows, again LPG-TD and SGPlan maintain their good scaling behavior from the simpler domain versions. To the other planners, there is a huge runtime performance gap. SGPlan is the only planner here that minimizes the number of actions. The makespan ratio LPG-TD.speed vs LPG-TD.quality is [1.00(1.15)2.01]. With compiled time windows, only SGPlan and CRIKEY participated. The former consistently solves the smaller instances up to a certain size, the latter fails on many of the small instances but can solve a few of the larger ones. Neither of the planners shows reasonable runtime performance compared to the explicit encoding of the domain. The number of actions ratio CRIKEY vs SGPlan is [1.00(1.14)1.36]. Our main motivation for including the Airport domain in IPC-4 was that we were able to model it quite realistically, and to generate quite realistic test instances - the only thing we could not model was the real optimization criterion (which asks to minimize the summed up travel time of all airplanes). So the good scaling results of, at least, the satisficing planners, are encouraging. Note, however, that in reality the optimization criterion is of crucial importance, and really the only reason for using a computer to control the traffic. 544 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEYPMEP LPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 TP4CPT HSPS_A (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN OPTOPPMEP TILSAPA 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 SGPLANCRIKEY (c) (d) Figure 2: Performance in temporal Airport, satisficing (a) and optimal (b). In temporal Airport with time windows, satisficing, explicit encoding (c) and compiled encoding (d). We remark that the domain is not overly difficult from the perspective of relaxed-plan based heuristic planners (using the wide-spread "ignore deletes" relaxation). Hoffmann (2005) shows that Airport instances can contain unrecognized dead ends (states from which the goal can't be reached, but from which there is a relaxed plan); but such states are not likely to occur very often. A dead end in Airport can only occur when two planes block each other, trying to get across the same segment in different directions. Due to the topology of the airports (with one-way roads), this can happen only in densely populated parking regions. In case for any two airplanes the respective paths to the goal position are disjoint, the length of a relaxed plan provides the exact distance to the nearest goal state. In that sense, the good runtime performance of the satisficing planners didn't come entirely unexpected to us, though we did not expect them to behave that well. Note that most of these planners, particularly FDD, SGPlan, and LPG-TD, do much more than/do things 545 Hoffmann & Edelkamp different from a standard heuristic search with goal distances estimated by relaxed plan length. Note also that Hoffmann (2005)'s results are specific to non-temporal domains. In the non-temporal domain version, we awarded 1st places to FDD (and FD) and SATPLAN; we awarded 2nd places to LPG-TD, SGPlan, Semsyn, and BFHSP. In the temporal version, we awarded 1st places to LPG-TD, SGPlan, and CPT; a 2nd place was awarded to TP4. In the domain version with explicit time windows, 1st places were awarded to LPG-TD and SGPlan. 5.4 Pipesworld The Pipesworld domain is a PDDL adaptation of an application domain dealing with complex problems that arise when transporting oil derivative products through a pipeline system. Note that, while there are many planning benchmarks dealing with variants of transportation problems, transporting oil derivatives through a pipeline system has a very different and characteristic kind of structure. The pipelines must be filled with liquid at all times, and if you push something into the pipe at one end, something possibly completely different comes out of it at the other end. As a result, the domain exhibits some interesting planning space characteristics and good plans sometimes require very tricky maneuvers. Additional difficulties that have to be dealt with are, for example, interface restrictions (different types of products interfere with each other in a pipe), tankage restrictions in areas (i.e., limited storage capacity defined for each product in the places that the pipe segments connect), and deadlines on the arrival time of products. In the form of the domain used in IPC-4, the product amounts dealt with are discrete in the sense that we assume a smallest product unit, called "batch". There were six different domain versions in IPC-4: notankage-nontemporal, notankage-temporal, tankagenontemporal, tankage-temporal, notankage-temporal-deadlines, and notankage-temporaldeadlines-compiled. All versions include interface restrictions. The versions with "tankage" in their name include tankage restrictions, modelled by a number of "tank slots" in each place in the network, where each slot can hold one batch (note that the slots introduce some additional symmetry into the problem; we get back to this below). In the versions with "temporal" in their name, actions take (different amounts of) time. The versions with "deadlines" in their name include deadlines on the arrival of the goal batches, modelled by timed initial literals respectively their compilation to standard PDDL2.1. None of the encodings uses any ADL constructs, so of each domain version there is just one (STRIPS) formulation. The IPC-4 example instances were generated based on five scaling network topologies. The smallest network topology has 3 areas connected by 2 pipes; the largest network topology has 5 places connected by 5 pipes. For each network we generated 10 instances, with growing numbers of batches, and of batches with a goal location. So altogether we had 50 instances per domain version, numbered consecutively. Between domain versions, the instances were, as much as possible, transferred by adding/removing constructs. E.g., the instances in tankage-nontemporal (tankage-temporal) are exactly the same as those in notankage-nontemporal (notankage-temporal) except that tankage restrictions were added. We do not include graphs for the domain versions featuring deadlines. In the version with explicit deadlines (modelled by timed initial literals), only LPG-TD and Tilsapa 546 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDMACRO-FF SGPLANCRIKEY FDD 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 FAPMARVIN ROADMAPPERPMEP YAHSPLPG-3 (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 BFHSPTP4 CPTOPTIPLAN SATPLANSEMSYN (c) Figure 3: Non-temporal Pipesworld, no tankage constraints, satisficing (a) and (b), optimal (c). participated, of which LPG-TD scaled up to middle-size instances; Tilsapa solved only a few of the smallest instances. In the domain version with compiled deadlines, only CPT participated, solving a few small instances. Figure 3 shows the results for the non-temporal domain version without tankage restrictions. Parts (a) and (b) contain the results for satisficing planners in the respective non-temporal domain version. We observe that YAHSP and SGPlan are the only planners that can solve all instances. YAHSP is a lot faster in most instances; but it finds excessively long plans: the ratio YAHSP vs SGPlan is [0.38(1.77)14.04], where in the maximum case SGPlan needs 72 actions, YAHSP 1,011. For the other planners, plan quality does not vary much, with the exception of Marvin, whose makespan ratio vs the optimal CPT is [0.88(1.60)2.19]. LPG-TD and FDD both solve all but a few of the largest instances. Note that, as in Airport, the IPC-3 version of LPG is outperformed dramatically by the best IPC-4 planners. 547 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEYPMEP LPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 TP4CPT HSPS_A (a) (b) Figure 4: Temporal Pipesworld without tankage constraints, satisficing (a), optimal (b). Part (c) of Figure 3 shows the results for the optimal planners. The runtime curves are extremely similar, nearing indistinguishable. Note that Optiplan is the only optimal planner that solves an instance with a very large size parameter. However, in Pipesworld, like in Airport, due to the domain combinatorics, planners are likely to find it easier to solve a large network with little traffic, than a small network with a lot of traffic.10 The other optimal planners here may just not have tried to run their planner on larger instances when it already failed on the smaller ones - we explicitly advised people to save machine workload by not insisting on spending half an hour runtime on instances that were probably infeasible for their planners anyway. In cases like this one here, this admittedly is a potentially misleading guideline. In parts (a) and (b) of Figure 4, we display the results in the temporal domain version without tankage restrictions. Part (a) shows a clear win of SGPlan over the other planners. We find it particularly remarkable that, as before in Airport, SGPlan is just as good in the temporal domain version as in the nontemporal one. Again, LPG-3 is outperformed by far. As for the optimal planners in part (b) of the figure, CPT scales best; TP4 and HSP*a scale similarly. The only planners minimizing the number of actions here are CRIKEY and SGPlan; the ratio is [0.35(1.24)5.67], showing quite some variance with a slight mean advantage for SGPlan. Of the planners minimizing makespan, LPG-TD and P-MEP sometimes return long plans; in one instance (number 2), LPG-TD's plan is extremely long (makespan 432). More precisely, the ratio LPG-TD.speed vs the optimal CPT is [1.00(4.55)36.00]; the ratio P-MEP vs CPT is [1.19(2.25)4.59]; the ratio LPG-TD.quality vs CPT is [0.73(1.05)1.62] (i.e., in particular the peak at instance 2 disappears for LPG-TD.quality). Note that, strangely at first sight, sometimes LPG-TD finds better plans here than the optimal CPT. 10. On the other hand, note at this point that the networks in Pipesworld do by far not grow as much as in Airport, where they grow from microscopic toy airports for the smallest instances to a real-world airport for the largest instances. As said, the smallest network in Pipesworld has 3 areas connected by 2 pipes, and the largest network has 5 places connected by 5 pipes. 548 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDMACRO-FF SGPLANCRIKEY FDD 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 FAPMARVIN ROADMAPPERYAHSP LPG-3 (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 BFHSPCPT OPTIPLANSATPLAN (c) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEYLPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 TP4CPT HSPS_A (d) (e) Figure 5: Pipesworld with tankage constraints: non-temporal satisficing (a) and (b), optimal (c), temporal satisficing (d), optimal (e). This is due to the somewhat simpler model of durative actions that CPT uses (Vidal & Geffner, 2004), making no distinction between the start and end time points of actions. 549 Hoffmann & Edelkamp In Figure 5, results are displayed for the two domain versions that do feature tankage restrictions. Very generally, we see that the planners have much more trouble with these domain versions than with their less constrained counterparts. In the non-temporal domain version with tankage constraints, for the satisficing planners in parts (a) and (b) we observe that, as before, YAHSP is most efficient, solving more instances than any of the other planners. Again, this efficiency is bought at the cost of overlong plans: the ratio YAHSP vs. SGPlan is [0.66(1.89)6.14], where in the maximum case SGPlan needs 64 actions, YAHSP 393. For the other planners, plan quality does not vary as much (e.g. the ratio FDD vs SGPlan is [0.56(0.96)1.38]. Of the optimal planners displayed in part (c), same non-temporal domain version, SATPLAN and Optiplan scale a little better than the other planners, solving two and three more instances, respectively. SATPLAN tends to be faster than Optiplan in the solved cases. Part (d) displays the results for the satisficing planners in the respective temporal domain version. SGPlan clearly scales best, once again keeping its performance from the corresponding non-temporal domain version. LPG-TD is followed relatively closely by CRIKEY but can solve some more instances. Regarding plan quality, CRIKEY and SGPlan minimize plan length, and the ratio CRIKEY vs SGPlan is [0.62(1.37)2.54]. For makespan, LPG-TD again has a peak in instance 2; the ratio LPG-TD.speed vs CPT is [1.00(2.55)7.27], maximum 160 vs 22 units; the ratio LPG-TD.quality vs CPT is [0.86(0.95)1.18] (LPG-TD.quality's makespan in instance 2 is 26). Note that, as above, LPG-TD.quality can find better plans than CPT due to the somewhat simpler model of durative actions used in CPT. In the optimal track, TP4 performs a little better, in terms of runtime, than the similar-performing CPT and HSP*a; all the planners solve only a few of the smallest instances. We remark that we were rather surprised by the relatively good scaling behavior of some planners, particularly of the fastest satisficing planners in the domain versions without tankage restrictions. Hoffmann (2005) shows that, in Pipesworld, there can be arbitrarily deep local minima under relaxed plan distances (i.e., the distance from a local minimum to a better region of the state space may be arbitrarily large). No particularly unusual constructs are needed to provoke local minima - for example, a cycle of pipe segments, as also occurs in the network topologies underlying the IPC-4 instances, suffices. Not less importantly, in the limited experiments we did during testing, FF (Hoffmann & Nebel, 2001) and Mips (Edelkamp, 2003c) generally scaled much worse than, e.g., YAHSP and SGPlan do as shown above. As for the domain versions that do feature tankage restrictions, these are harder than their counterparts. There are two important differences between the domain versions with/without tankage restrictions. First, the problem is more constrained (remember that, modulo the tankage constraints, the IPC-4 instances are identical in both respective versions). Second, the tank slots, that model the restrictions, introduce additional symmetry into the problem: a planner can now choose into which (free) slot to insert a batch. This choice does not make a difference, but enlarges the state space exponentially in the number of such choices (to the basis of the number of tank slots). We considered the option to use a "counter" encoding instead, to avoid the symmetry: one could count the product amounts in the areas, and impose the tankage restrictions based on the counter values. We decided against this because symmetry is a challenging aspect of a benchmark. 550 The Deterministic Part of IPC-4: An Overview In domain version notankage-nontemporal, we awarded 1st places to YAHSP and SGPlan; we awarded 2nd places to LPG-TD and FDD. In version notankage-temporal, we awarded 1st places to SGPlan and CPT; we awarded 2nd places to LPG-TD, TP4, and HSP*a. In version tankage-nontemporal, we awarded 1st places to YAHSP and FDD; we awarded 2nd places to SGPlan, SATPLAN, and Optiplan. In version tankage-temporal, we awarded a 1st place to SGPlan; we awarded 2nd places to LPG-TD and TP4. In version notankage-temporal-deadlines, we awarded a 1st place to LPG-TD. 5.5 Promela Promela is the input language of the model checker SPIN (Holzmann, 2004), used for specifying communication protocols. Communication protocols are distributed software systems, and many implementation bugs can arise, like deadlocks, failed assertions, and global invariance violations. The model checking problem (Clarke, Grumberg, & Peled, 2000) is to find those errors by returning a counter-example, or to verify correctness by a complete exploration of the underlying state-space. Edelkamp (2003b) developed an automatic translation of this problem, i.e., of the Promela language, into (non-temporal) PDDL, making it possible to apply state-of-the-art planners without any modification. For IPC-4, two relatively simple communication protocols were selected as benchmarks - toy problems from the Model-Checking area. One is the well-known Dining-Philosophers protocol, the other is a larger protocol called Optical-Telegraph. The main point of using the domain, and these protocols, in IPC-4 was to promote the connection between Planning and Model-Checking, and to test how state-of-the-art planners scale to the basic problems in the other area. The IPC-4 instances exclusively require to find deadlocks in the specified protocols - states in which no more transitions are possible. The rules that detect whether or not a given state is a deadlock are most naturally modelled by derived predicates, with complex ADL formulas in the rule bodies. To enable broad participation, we provided compilations of derived predicates into additional actions, and of ADL into STRIPS. Edelkamp's original translation of Promela into PDDL also makes use of numeric variables. For the non-numeric planners, the translation was adapted to use propositional variables only - for the relatively simple Dining-Philosophers and Optical-Telegraph protocols, this was possible. Precisely, the domain versions and formulations were the following. First, the versions were split over the modelled protocol, Dining-Philosophers or Optical-Telegraph. Second, the versions were split over the used language: with/without numeric variables, and with/without derived predicates.11 So, all in all, we had 8 domain versions. Each of them used ADL constructs. In those 4 versions without numeric variables, we also provided STRIPS formulations, obtained from the ADL formulations with automatic compilation software based on FF's pre-processor (Hoffmann & Nebel, 2001), producing fully grounded encodings. All Promela domain versions are non-temporal. Here, we show plots only for the results obtained in the non-numeric domain versions. In the domain versions using both numeric variables and derived predicates, no planner 11. We split with/without numeric variables into different domain versions, rather than formulations, to encourage the use of numeric planning techniques. We split with/without derived predicates into different versions since, through the modelling of derived predicates as new actions, the plans become longer. 551 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 LPG-TDMACRO-FF SGPLANCRIKEY FAPPMEP YAHSP 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 BFHSPTP4 CPTHSPS_A OPTIPLANSATPLAN SEMSYN (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 LPG-TDSGPLAN FDDMARVIN (c) Figure 6: Performance in Promela/Dining Philosophers. Encoding without derived predicates, satisficing (a), optimal (b). Encoding with derived predicates, satisficing (c). participated. In the numeric version of Dining-Philosophers (without derived predicates), only SGPlan and P-MEP participated. SGPlan solved all instances while P-MEP solved only a few of the smallest ones. In the numeric version of Optical-Telegraph (without derived predicates), only SGPlan participated, solving some relatively small instances. Parts (a) and (b) of Figure 6 show the results for Dining-Philosophers, without derived predicates, i.e. with additional actions to derive the deadlocks. In the the satisficing track, YAHSP and SGPlan clearly show the best performance. The other planners all lag several orders of magnitude behind. In the optimal track, SATPLAN clearly scales best, followed by Optiplan. Observe that, in difference to what we have seen in Airport and Pipesworld, the optimal planners are just as efficient as the satisficing ones. This is evident from the plots. Most particularly, SATPLAN and Optiplan solve just as many instances, up to 30 philosophers (instance number x features x + 1 philosophers), like YAHSP and 552 The Deterministic Part of IPC-4: An Overview SGPlan.12 SATPLAN even does so in comparable time. Such a competitivity of optimal planners with satisficing ones has not been seen in any test suite used in any of the last three competitions.13 The efficiency of SATPLAN and Optiplan in Dining-Philosophers is probably due the fact that the needed number of parallel time steps is constantly 11 across all instances (see below). In a (standard) planning graph (Blum & Furst, 1995, 1997), the goals are first reached after 7 steps; so only 4 unsuccessful iterations are made before a plan is found. Regarding plan quality, there is one group of planners trying to minimize the number of actions, and one group trying to minimize the makespan (the number of parallel action steps). The results clearly point to an important difference between this domain and the other IPC-4 domains. Namely, there is only a single scaling parameter n - the number of philosophers - and only a single instance per value of n. The optimal number of actions is a linear function of n, precisely 11n: basically, one blocks all philosophers in sequence, taking 11 steps for each. The optimal makespan is, as mentioned above, constantly 11: one can block all philosophers in parallel. The only sub-optimal plan found by any planner minimizing the number of actions is that of CRIKEY for n = 8, containing 104 instead of 88 actions. Of the planners minimizing makespan, only P-MEP finds sub-optimal plans: it solves only the smallest four instances, n = 2, 3, 4, 5, with a makespan of 19, 25, 30, and 36, respectively. Figure 6 (c) shows the results in the domain version that uses derived predicates to detect the deadlock situations. No optimal planner could handle derived predicates which is why there are no results for this. Of the satisficing planners, SGPlan clearly shows the best performance. FDD and LPG-TD are roughly similar - note that, while LPG-TD is faster than FDD in most examples, the behavior of FDD in the largest instances indicates an advantage in scaling. There is a sudden large increase in LPG-TD's runtime, from the second largest instance (that is solved in 52 seconds) to the largest instance (that is solved in 1045 seconds). In difference to that, FDD solves the largest instance in almost the same time as the second largest one, taking 111 seconds instead of 110 seconds (in fact, FDD's runtime performance shows little variance and is pretty much a linear function in the instance size). We remark that the plans for the largest of these instances are huge: more than 400 steps long, see below. SGPlan generates these plans in little more than a single second CPU time. Regarding plan quality, in this domain version the optimal number of actions is 9n, and the optimal makespan is constantly 9. All planners except Marvin tried to minimize the number of actions. FD and SGPlan always find optimal plans. FDD, LPG-TD.speed, and LPG-TD.quality sometimes find slightly sub-optimal plans. Precisely, the ratio FDD vs FD is [1.00(1.08)1.44]; LPG-TD.speed vs FD is [1.00(1.10)2.19]; LPG-TD.quality vs FD is [1.00(1.03)1.24]. As for Marvin, the makespan of its plans is, roughly, linear in n; it always lies between 7n and 9n. 12. Importantly, these planners handle only STRIPS, and for compilation reasons there were STRIPS instances up to n = 30 only. So these planners solve all their respective test instances. We discuss this in a little more detail further below. 13. In the first IPC, in 1998, apart from the fact that the "heuristic search" satisficing planners were still in their infancy, Mystery and Mprime were used as benchmarks. These can't be solved particularly efficiently by any planner up to today, except, maybe, FD and FDD (Helmert, 2004). 553 Hoffmann & Edelkamp At this point, there is an important remark to be made about a detail regarding our compilation techniques in these domain versions. Observe the huge gap in planner performance between Figure 6 (a) and (b), and Figure 6 (c). As we have seen, the difference in plan length (makespan) is not very large - 11n (11) compared to 9n (9). Indeed, the apparent performance difference is due to a compilation detail, inherent in the IPC-4 instances, rather than to planner performance. In the version without derived predicates, we were able to compile only the instances with up to n = 30 philosophers from ADL into STRIPS. For larger values of n, the (fully-grounded) STRIPS representations became prohibitively large. In the version with derived predicates, the blow-up was a lot smaller, and we could compile all instances, with up to n = 49 philosophers. SGPlan, YAHSP, SATPLAN, and Optiplan can all handle only STRIPS. So, as mentioned above, in Figure 6 (a) and (b) these planners actually solve all instances in their respective test suite; they would probably scale up further when provided with STRIPS representations of instances with larger n values. 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 LPG-TDMACRO-FF SGPLANCRIKEY FAPPMEP YAHSP 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 BFHSPTP4 CPTHSPS_A OPTIPLANSATPLAN (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 LPG-TDSGPlan FDDMARVIN (c) Figure 7: Performance in Promela/Optical Telegraph. Encoding without derived predicates, satisficing (a), optimal (b). Encoding with derived predicates, satisficing (c). 554 The Deterministic Part of IPC-4: An Overview The Optical-Telegraph protocol is slightly more complex than the Dining-Philosopher example, i.e. it involves more complicated and indirect interactions between the communicating processes, leading to longer solutions. There is still just one scaling parameter, the number n of "telegraph station pairs", and a single instance per value of n. Each telegraph station pair is a pair of processes that goes through a rather complicated internal communication structure, enabling the exchange of data. The telegraph station pair shares with the outside world - i.e., with the other telegraph station pairs - two "control channels" that must be occupied as a prerequisite to the internal exchange of data. Thus, the role of the control channels is pretty similar to the role of the (shared) forks in the Dining-Philosopher example. Instance number x in IPC-4 features x + 1 telegraph station pairs. Figure 7 shows in parts (a) and (b) the performance of the satisficing and optimal planners, respectively, in the domain version using additional actions to derive deadlocks. In the group of satisficing planners, Macro-FF clearly performs best. Of the other satisficing planners, SGPlan is the most efficient. It is important to note that, here, we were able to compile only the instances up to n = 15 from ADL into STRIPS. So SGPlan and LPG-TD solve all their respective test instances. Of the optimal planners, SATPLAN and Optiplan solve the instances up to n = 14, i.e., they failed to solve the largest instance in the STRIPS test suite they attacked. SATPLAN is much faster than Optiplan. Each of the other optimal planners could solve only the smallest instance, n = 2. Figure 7 (c) shows that FDD, which handles the ADL formulation, is by far the most successful satisficing planner in the domain version encoding Optical-Telegraph with derived predicates for deadlock detection; of the other planners, SGPlan scales best, solving all instances in the STRIPS set, up to n = 20. Regarding plan quality, in Optical-Telegraph without derived predicates the optimal number of actions is 18n, and the optimal makespan is constantly 13: optimal sequential plans block all telegraph station pairs in sequence, taking 18 steps for each; in parallel plans, some simultaneous actions are possible within each telegraph station pair. In OpticalTelegraph with derived predicates, the optimal number of actions is 14n, and the optimal makespan is constantly 11. In the competition results, all planners returned the optimal plans in these test suites, in all cases - with a single exception. The plans found by Marvin in the version with derived predicates have makespan 14n in all solved cases. As we have seen, the results in Promela are, over all, quite different from, e.g., those we have seen in Airport and Pipesworld. There is only a single scaling parameter and a single instance per size, and optimal makespan is constant. This leads to rather smooth runtime and plan quality curves, as well as to an unusual competitivity of optimal planners with satisficing planners. The scalability of the planners both shows that current planners are able to efficiently solve the most basic Model-Checking benchmarks (Dining-Philosophers), and that they are not very efficient in solving more complex Model-Checking benchmarks (Optical-Telegraph). We remark that Hoffmann (2005) shows that there exist arbitrarily deep local minima under relaxed plan distances in Optical-Telegraph, but not in DiningPhilosophers, where there exists a large upper bound (31) on the number of actions needed to escape a local minimum. It is not clear, however, how much these theoretical results have to do with the planner performance observed above. The worst-cases observed by Hoffmann all occur in regions of the state space not entered by plans with optimal number of actions - as found by the IPC-4 participants, in most cases. 555 Hoffmann & Edelkamp In Dining-Philosophers without derived predicates, we awarded 1st places to YAHSP, SGPlan, and SATPLAN; we awarded a 2nd place to Optiplan. In Dining-Philosophers with derived predicates, we awarded a 1st place to SGPlan, and 2nd places to FDD and LPG-TD. In Optical-Telegraph without derived predicates, we awarded 1st places to Macro-FF and SATPLAN; we awarded 2nd places to SGPlan and Optiplan. In Optical-Telegraph with derived predicates, we awarded a 1st place to FDD, and a 2nd place to SGPlan. In the numeric version of Dining-Philosophers, we awarded a 1st place to SGPlan. 5.6 PSR PSR is short for Power Supply Restoration. The domain is a PDDL adaptation of an application domain investigated by Sylvie Thi'ebaux and other researchers (Thi'ebaux, Cordier, Jehl, & Krivine, 1996; Thi'ebaux & Cordier, 2001), which deals with reconfiguring a faulty power distribution system to resupply customers affected by the faults. In the original PSR problem, various numerical parameters such as breakdown costs and power margins need to be optimized, subject to power capacity constraints. Furthermore, the location of the faults and the current network configuration are only partially observable, which leads to a tradeoff between acting to resupply lines and acting to reduce uncertainty. In contrast, the version used for IPC-4 is set up as a pure goal-achievement problem, numerical aspects are ignored, and total observability is assumed. Temporality is not a significant aspect even in the original application, and the IPC-4 domain is non-temporal. We used four domain versions of PSR in IPC-4. Primarily, these versions differ by the size of the problem instances encoded. The instance size determined in what languages we were able to formulate the domain version. The domain versions are named 1. large, 2. middle, 3. middle-compiled, and 4. small. Version 1 has the single formulation adl-derivedpredicates: in the most natural formulation, the domain comes with derived predicates to model the flow of electricity through the network, and with ADL formulas to express the necessary conditions on the status of the network connectors. Version 2 has the formulations adl-derivedpredicates, simpleadl-derivedpredicates, and strips-derivedpredicates. Version 3 has the single formulation adl, and version 4 has the single formulation strips. As indicated, the formulation names simply give the language used. In version 2, ADL constructs were compiled away to obtain the simpler formulations. In version 3, derived predicates were compiled away by introducing additional artificial actions; due to the increase in plan length (which we discuss in some detail further below), we turned this into a separate domain version, rather than a formulation. As for the strips domain version, to enable encoding of reasonably-sized instances for this we adopted a different fully-grounded encoding inspired by the work of Bertoli et al. (2002). The encoding is generated from a description of the problem instance by a tool performing some of the reasoning devoted to the planner under the other domain versions. Still we were only able to formulate comparatively small instances in pure STRIPS. Starting with the performance in the smallest instances, we depict the performance of satisficing and optimal planners in the PSR STRIPS domain in Figure 8. Both result graphs are divided into two because they are completely unreadable otherwise. Most planners show a lot of variance in this domain, blurring the performance differences (observable) between the individual systems. Still it is possible to identify systems that behave better than others. 556 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEY 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 FDDMARVIN YAHSPLPG-3 (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 BFHSPTP4 CPTHSPS_A 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 OPTIPLANSATPLAN SEMSYN (c) (d) Figure 8: PSR small, satisficing (a) and (b), optimal (c) and (d). Of the satisficing systems, only FDD can solve all the 50 instances; it does so consistently fast (the results for FD are almost identical). Of the other planners, LPG-TD, SGPlan, and YAHSP have the best success ratio: they solve 49, 47, and 48 instances, respectively; CRIKEY solves 29 instances, Marvin 41. As for the optimal systems, here the only system solving the entire test suite is SATPLAN. BFHSP solved 48 instances, CPT 44, HSP*a 44, Optiplan 29, Semsyn 40, TP4 38. For plan quality, once again there are groups of planners trying to minimize plan length (number of actions), and makespan. In the former group, we take as a performance measure the plans found by BFHSP, which are optimal in that sense, and which, as said above, we have for all but 2 of the 50 instances. The ratio CRIKEY vs BFHSP is [1.00(1.72)3.44]; the ratio FDD vs BFHSP is [1.00(1.02)1.52]; the ratio LPG-TD.speed vs BFHSP is [1.00(5.52)12.70]; the ratio LPG-TD.quality vs BFHSP is [1.00(1.82)8.32]; the ratio SGPlan vs BFHSP is [1.00(1.01)1.24]; the ratio YAHSP vs BFHSP is [1.00(1.00)1.05]. In the planner group minimizing makespan, all planners except Marvin are optimal. The ratio Marvin vs SATPLAN is [1.00(1.28)2.07]. 557 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN FDDMARVIN 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 SGPLANFDD MARVIN (a) (b) Figure 9: PSR, middle (a) and large (b), satisficing planners. Apart from the observations to be made within the groups of satisficing respectively optimal planners, there is something to be said here on the relationship between these two groups. Like in Dining-Philosophers, we have the rather unusual situation that the optimal planners are just as efficient as the satisficing ones. Indeed, by solving all instances, SATPLAN is superior to most of the satisficing planners. We remark at this point that Hoffmann (2005) shows the existence of arbitrarily deep local minima in PSR, i.e., regions where it takes arbitrarily many step to escape a local minimum under relaxed plan distances. For example, local minima arise naturally because a relaxed plan is able to supply and not supply a line at the same time, and thereby does not have to make the crucial distinction between faulty and non-faulty lines. Now, these observations hold for the original domain formulation, with derived predicates and ADL constructs. As said, the IPC-4 STRIPS formulation of PSR is obtained from that by a combination of complicated pre-processing machines. These are not likely to make the real structure of the domain more amenable to relaxed plan distances. While the satisficing planners participating in PSR small are by no means exclusively dependent on relaxed plan distances, for most of them these distances do form an important part of the search heuristics. Figure 9 gives the results for the PSR domain versions, middle and large, with larger instances, encoded with ADL constructs and/or derived predicates. No optimal planner could handle these language constructs, so unfortunately the above comparison between the two groups can not be continued. In the middle sized instances, FDD, SGPlan, and LPG-TD all scale through the entire test suite. FDD indicates better scaling behavior in the largest instances. In the domain version with large instances (available only in a formulation using ADL), FDD indeed shows that it outperforms the other planners (at least, SGPlan, that also participates here) by far. The data for FD is almost identical to that for FDD. Regarding plan quality, the only participating planner in each of middle and large that tries to minimize makespan is Marvin, so we have no basis for a comparison. Of the other planners, SGPlan generally finds the plans with the smallest number of actions. Precisely, in the middle version the plan quality ratios are as follows. The ratio FDD vs 558 The Deterministic Part of IPC-4: An Overview SGPlan is [0.67(1.23)2.33] (FD: [0.67(1.25)2.85]). The ratio LPG-TD.speed vs SGPlan is [0.67(1.82)7.27]; the maximum case is instance number 49, where LPG-TD.speed takes 80 steps and SGPlan 11. The ratio LPG-TD.quality vs SGPlan is [0.60(1.08)7.27], with the same maximum case. In the large version, only FD and FDD solve a considerable number of instances; the ratio FD vs FDD is [0.60(1.01)1.30]. In the PSR domain version with middle-size instances and derived predicates compiled into ADL, i.e., in domain version middle-compiled, only Macro-FF and SGPlan participated. Macro-FF scaled relatively well, solving 32 instances up to instance number 48. SGPlan solved only 14 instances up to instance number 19. Both planners try to minimize the number of actions, and the ratio SGPlan vs Macro-FF is [0.51(1.28)1.91]. It is interesting to observe that, in PSR, when compiling derived predicates away, the plan length increases much more than what we have seen in Section 5.5 for DiningPhilosophers and Optical-Telegraph. In the latter, the derived predicates just replace two action applications per process, detecting that the respective process is blocked. In PSR, however, as said the derived predicates model the flow of electricity through the network, i.e., they encode the transitive closure of the underlying graph. The number of additional actions needed to simulate the latter grows, of course, with the size of the graph. This phenomenon can be observed in the IPC-4 plan length data. The PSR middle-compiled instances are identical to the middle instances, except for the compilation of derived predicates. Comparing the plan quality of Macro-FF in middle-compiled with that of SGPlan in middle, we obtain the remarkably high ratio values [7.17(12.53)25.71]. The maximum case is in instance number 48 - the largest instance solved by Macro-FF - where Macro-FF takes 180 steps to solve the compiled instance, while SGPlan takes only 7 steps to solve the original instance.14 In domain version small, we awarded 1st places to FDD (and FD), and SATPLAN; we awarded 2nd places to LPG-TD, SGPlan, YAHSP, and BFHSP. In domain version middle, we awarded a 1st place to FDD (and FD); we awarded 2nd places to LPG-TD and SGPlan. In domain version middle-compiled, we awarded a 1st place to Macro-FF. In domain version large, we awarded a 1st place to FDD (and FD). 5.7 Satellite The Satellite domain was introduced in IPC-3 by Long and Fox (2003). It is motivated by a NASA space application: a number of satellites has to take images of a number of spatial phenomena, obeying constraints such as data storage space and fuel usage. In IPC-3, there were the domain versions Strips, Numeric, Time (action durations are expressions in static variables), and Complex (durations and numerics, i.e. the "union" of Numeric and Time). The numeric variables transport the more complex problem constraints, regarding data capacity and fuel usage. For IPC-4, the domain was made a little more realistic by additionally introducing time windows for the sending of the image data to earth, i.e. to antennas that are visible for satellites only during certain periods of time. We added the new domain versions Time14. While Macro-FF is not an optimal planner, and the optimal plan lengths for the middle-compiled instances are not known, it seems highly unlikely that these observations are only due to overlong plans found by Macro-FF. 559 Hoffmann & Edelkamp timewindows, Time-timewindows-compiled, Complex-timewindows, and Complex-timewindowscompiled, by introducing time windows, explicit and compiled, into the IPC-3 Time and Complex versions, respectively.15 None of the domain versions uses ADL constructs, so of all versions there was only a single (STRIPS) formulation. The instances were, as much as possible, taken from the original IPC-3 instance suites. Precisely, the Strips, Numeric, Time, and Complex versions each contained the 20 instances posed in IPC-3 to the fully-automatic planners, plus the 16 instances posed in IPC-3 to the hand-tailored planners. For Timetimewindows and Time-timewindows-compiled, we extended the 36 instances from Time with the time windows. Similarly, for Complex-timewindows and Complex-timewindowscompiled, we extended the 36 instances from Complex with the time windows. That is, in the newly added domain versions the sole difference to the previous instances lies in the time windows. Let us first consider the results in the simpler versions of the Satellite domain. Figure 10 shows the results for the Strips and Time versions. To many of the satisficing planners, the Strips version does not pose a serious problem, see Figure 10 (a). Macro-FF and YAHSP show the best runtime behavior, followed closely by LPG-TD and FDD (as well as FD). As for the optimal planners in Figure 10 (b), none of them scales up very far. SATPLAN is most efficient, solving 11 instances; CPT and Semsyn each solve 9 instances, Optiplan solves 8, BFHSP solves only 6. In the Time version, LPG-TD and SGPlan behave similarly up to instance number 30, but LPG-TD solves two more, larger, instances. The Time optimal planners, see Figure 10 (d), are clearly headed by CPT. Regarding the efficiency of the satisficing planners in the Strips test suite, we remark that these instances were solved quite efficiently at IPC-3 already, e.g. by FF and LPG-3. Further, Hoffmann (2005) shows that, in this domain, under relaxed plan distance, from any non-goal state, one can reach a state with strictly smaller heuristic value within at most 5 steps. In that sense, the results depicted in Figure 10 (a) didn't come as a surprise to us. Let us consider plan quality in Strips and Time. In both domain versions, as before the addressed optimization criteria are plan length, and makespan. In Strips, the plan quality behavior of all the planners trying to minimize plan length is rather similar. The overall shortest plans are found by LPG-TD.quality. Precisely, the ratio FDD vs LPG-TD.quality is [0.96(1.19)1.48]; the ratio FD vs LPG-TD.quality is [0.96(1.20)1.53]; the ratio LPGTD.speed vs LPG-TD.quality is [1.00(1.12)1.31]; the ratio Macro-FF vs LPG-TD.quality is [0.95(1.03)1.17]; the ratio Roadmapper vs LPG-TD.quality is [1.00(1.35)1.71]; the ratio SGPlan vs LPG-TD.quality is [0.99(1.07)1.70]; the ratio YAHSP vs LPG-TD.quality is [0.97(1.22)1.93]. Of the planners trying to optimize makespan in the Strips version, only Marvin and P-MEP are satisficing; Marvin is the only planner that scales to the larger instances. The ratio P-MEP vs SATPLAN (which solves a superset of the instances solved by P-MEP) is [1.01(2.22)3.03]. The ratio Marvin vs SATPLAN is [1.08(2.38)4.00]. In the Time version, the only planner minimizing plan length (i.e., the number of actions) is SGPlan, so there is no basis for comparison. The only satisficing planners trying 15. In the new domain versions derived from Complex, we also introduced utilities for the time window inside which an image is sent to earth. For each image, the utility is either the same for all windows, or it decreases monotonically with the start time of the window, or it is random within a certain interval. Each image was put randomly into one of these classes, and the optimization requirement is to minimize a linear combination of makespan, fuel usage, and summed up negated image utility. 560 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDMACRO-FF SGPLANFDD MARVINPMEP ROADMAPPERYAHSP LPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 BFHSPCPT OPTIPLANSATPLAN SEMSYN (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDSGPLAN PMEPLPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 TP4CPT HSPS_A (c) (d) Figure 10: Satellite; Strips satisficing (a), optimal (b), and Time satisficing (c), optimal (d). to minimize makespan are P-MEP and LPG-TD. In the small instances solved by CPT (actually, a superset of those solved by P-MEP), the ratio P-MEP vs CPT is [1.10(3.71)6.49]. The ratio LPG-TD.speed vs CPT is [1.33(3.37)5.90]; the ratio LPG-TD.quality vs CPT is [0.85(1.24)1.86]. Note that, like we have seen in Pipesworld before, sometimes LPG-TD finds better plans here than the optimal CPT. As said, this is due to the somewhat simpler model of durative actions that CPT uses (Vidal & Geffner, 2004), making no distinction between the start and end time points of actions. Figure 11 (a) shows the results in Satellite Numeric, together for all participating planners - the only optimal planner that participated here was Semsyn. SGPlan and LPG-TD scale best; they solve the same number of instances, but SGPlan solves some larger ones and is at least an order of magnitude faster in those instances solved by both. Regarding plan quality, SGPlan was the only planner here trying to minimize plan length. The other planners tried to minimize the metric value of the plan, i.e., the quality metric specified in the instance files, which is fuel usage. The best plans in this respect are found by LPG561 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDSGPLAN PMEPSEMSYN LPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDSGPLAN (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDSGPLAN PMEPLPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 LPG-TDSGPLAN TILSAPA (c) (d) Figure 11: Satellite; Numeric (a), Time-timewindows (b), satisficing Complex (c), Complex-timewindows (d). TD.quality; it is unclear to us why Semsyn was marked to optimize the metric value here, and why it does not find the optimal plans. The ratio LPG-TD.speed vs LPG-TD.quality is [1.00(2.26)4.69]; the ratio P-MEP vs LPG-TD.quality is [1.01(2.16)3.03]; the ratio Semsyn vs LPG-TD.quality is [1.00(1.28)1.58]. In Time-timewindows, no optimal planner could participate due to the timed initial literals. Of the satisficing planners, see Figure 11 (b), only SGPlan and LPG-TD participated, which both scaled relatively well, with a clear advantage for SGPlan. Regarding plan length, SGPlan minimizes the number of actions, and LPG-TD the makespan. The ratio LPG-TD.speed vs LPG-TD.quality is [1.00(1.22)1.51]. In Time-timewindows-compiled, only SGPlan and CPT participated. SGPlan largely maintained its performance from the Time-timewindows version, CPT could solve only 3 of the smallest instances. Plan quality can't be compared due to the different criteria (plan length and makespan) that are minimized. 562 The Deterministic Part of IPC-4: An Overview Figure 11 (c) shows the performance of the satisficing planners in Satellite Complex. SGPlan and LPG-TD scale well, in particular much better than the only other competitor, P-MEP. In the largest three instances, SGPlan shows a clear runtime advantage over LPGTD. In the optimal track, only TP4 and HSP*a competed here. Both solved the same four very small instances, numbers 1, 2, 3, and 5, in almost the same runtime. SGPlan is the only planner minimizing the number of actions. Of the other planners, which all try to minimize makespan, only LPG-TD scales up to large instances. The ratio LPG-TD.speed vs LPGTD.quality is [1.01(2.76)4.71]. The ratio LPG-TD.speed vs TP4 is [1.81(2.09)2.49]; the ratio LPG-TD.quality vs TP4 is [0.93(1.07)1.19]; the ratio P-MEP vs TP4 is [1.27(2.25)3.32]. As above for CPT, the better plan (in one case) found by LPG-TD.quality is due to the somewhat simpler action model used in TP4 (Haslum & Geffner, 2001). Figure 11 (d) shows the performance of all planners in Complex-timewindows - as above, due to the timed initial literals no optimal planner could compete here. SGPlan scales clearly best, followed by LPG-TD; Tilsapa solves only the 3 smallest instances. Of these 3 participating planners, each one minimizes a different quality criterion: number of actions for SGPlan, makespan for Tilsapa, and metric value - the aforementioned linear combination of makespan, fuel usage, and summed up negated image utility - for LPG-TD. So the only useful comparison is that between LPG-TD.speed and LPG-TD.quality, where the ratio is [1.00(2.33)7.05].16 In Complex-timewindows-compiled, the only participating planner was SGPlan. It maintained its good scalability from the Complex domain version with explicit time windows. In domain version Strips, we awarded 1st places to Macro-FF, YAHSP, and SATPLAN; we awarded 2nd places to FDD (and FD), LPG-TD, CPT, Optiplan, and Semsyn. In domain version Time, we awarded 1st places to LPG-TD and CPT; we awarded a 2nd place to SGPlan. In each of the domain versions Numeric, Time-timewindows, Complex, and Complex-timewindows, we awarded a 1st place to SGPlan and a 2nd place to LPG-TD. 5.8 Settlers The Settlers domain was also introduced in IPC-3 (Long & Fox, 2003). It features just a single domain version, which makes extensive use of numeric variables. These variables carry most of the domain semantics, which is about building up an infrastructure in an unsettled area, involving the building of houses, railway tracks, sawmills, etc. In IPC-3, no planner was able to deal with the domain in an efficient way - the best IPC-3 planner in Settlers, Metric-FF (Hoffmann, 2003), solved only the smallest six instances of the test suite. For these reasons, we included the domain into IPC-4 as a challenge for the numeric planners. We used the exact same domain file and example instances as in IPC-3, except that we compiled away some universally quantified preconditions to improve accessibility for planners. The quantifiers were not nested, and ranged over a fixed set of domain constants, so they could easily be replaced by conjunctions of atoms. 16. Due to the negated image utility, the quality values here can be negative. For LPG-TD.speed and LPGTD.quality here, this happened in 5 instances. We skipped these when computing the given ratio values, since putting positive and negative values together doesn't make much sense - with negative quality, the number with larger absolute value represents the better plan. It would probably have been better to define, in this domain version, an image penalty instead of an image utility, and thus obtain strictly positive action "costs". 563 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 2 4 6 8 10 12 14 16 18 20 LPG-TDSGPLAN SEMSYN Figure 12: Performance of all planners in Settlers. Figure 12 shows the runtime results obtained in IPC-4. The efficiency increase compared to IPC-3 is, obviously, dramatic: SGPlan solves every instance within less than 15 seconds (instance number 8 is unsolvable).17 LPG-TD solves 13 instances out of the set, Semsyn solves 3 (numbers 1,2, and 5). We remark that the solutions for these tasks, as returned by SGPlan, are huge, up to more than 800 actions in the largest tasks. On the one hand, this once again demonstrates SGPlan's capability to find extremely long plans extremely quickly. On the other hand, SGPlan's plans in Settlers might be unnecessarily long, to some extent. In the largest instance solved by Semsyn, number 5, Semsyn finds a plan with 94 actions, while the plan found by SGPlan has 264 actions. In the largest instance solved by LPG-TD, number 17, LPG-TD.quality takes 473 actions while SGPlan takes 552. LPG-TD minimizes the metric value of the plans, a linear combination of the invested labor, the resource use, and the caused pollution. The ratio LPG-TD.speed vs LPG-TD.quality is [1.00(1.21)3.50]. We awarded a 1st place to SGPlan and a 2nd place to LPG-TD. 5.9 UMTS UMTS applications require comparatively much time to be started on a hand-held device, since they have to communicate with the network several times. If there is more than one application that has been called, this yields a true bottleneck for the user. Therefore, the applications set-up is divided into several parts to allow different set-up modules to work concurrently. The task in the domain is to provide a good schedule - minimizing the used time - for setting up timed applications, respecting the dependencies among them. In IPC-4, there were the six domain versions UMTS, UMTS-timewindows, UMTStimewindows-compiled, UMTS-flaw, UMTS-flaw-timewindows, and UMTS-flaw-timewindowscompiled. All of these domain versions are temporal, and make use of numeric variables to model the properties of the applications to be set up. ADL constructs are not used. UMTS 17. We remark that instance 8 is trivially unsolvable. One of the goals can only be achieved by actions having a static precondition that is false in the initial state. This can be detected by simple reachability analyses like, e.g., planning graphs, even without taking account of delete lists. 564 The Deterministic Part of IPC-4: An Overview is our standard model of the domain. In UMTS-timewindows there are additional time windows regarding the executability of set up actions, encoded with timed initial literals. In UMTS-timewindows-compiled the same time windows are compiled into artificial constructs. The remaining three domain versions result, as their names suggest, from adding a flaw construction to their respective counterparts. The flaw construction is practically motivated. It consists of an extra action that has an important sub-goal as its add effect, but that deletes another fact that can't be re-achieved. So the flaw action can't be used in a plan, but it can be used in relaxed plans, i.e. when ignoring the delete effects. In particular, adding an important sub-goal, the flawed action provides a kind of "short-cut" for relaxed plans, where the short-cut does not work in reality. This can lead to overly optimistic heuristic values. Thus the flaw may confuse the heuristic functions of relaxed-plan based heuristic planners. We used the flawed domain versions in IPC-4 to see whether the latter would be the case. In the IPC-4 test suites, all instances, irrespective of their number/size, contain 10 applications. The main scaling parameter is the number of applications that must actually be set up. For IPC-4 instances number 1 . . . 5, a single application must be set up; for instances number 6 . . . 10, two applications must be set up, and so on, i.e., the number of needed applications is bx/5c where x is the index of the instance. Figure 13 (a) and (b) shows the IPC-4 performance in the basic domain version. Obviously, SGPlan and LPG-TD have no difficulty at all with the domain - they solve every instance within split seconds. CRIKEY takes more time, but also scales up nicely. As for the optimal planners, only TP4 and HSP*a were able to handle the domain syntax, due to the combination of numeric variables and action durations. As Figure 13 (b) shows they scaled relatively similarly, with a slight advantage for HSP*a. Note that the two planners scaled better than P-MEP and LPG-3 in the satisficing track. We remark at this point that UMTS is mainly intended as a benchmark for optimal planners minimizing makespan. The domain is a pure scheduling problem by nature. As in most scheduling problems, it is trivial to find some plan - for example, one can simply schedule all applications in sequence.18 The only point of the domain, and, indeed, of using a computer to solve it, is to provide good schedules, that is, schedules with the smallest possible execution time, corresponding to the makespan of the plan. That said, we observe that the satisficing planners in IPC-4 were quite good at finding near-optimal plans in UMTS. In fact, as we will see in detail in the following, the only planner finding highly non-optimal plans was LPG-3. Let's consider the basic domain version treated in Figure 13 (a) and (b). Two of the participating planners, CRIKEY and SGPlan, try to minimize the number of actions (i.e., the wrong optimization criterion). Their data is identical, i.e. the plan "lengths" are the same for all instances. Precisely, both planners find, for each instance number x, a plan with bx/5c*8 actions in it. This is, in fact, the optimal (smallest possible) number of actions - remember what we said above about the scaling in the IPC-4 test suites. The participating planners trying to minimize makespan are LPG-3, LPG-TD, P-MEP, HSP*a, and TP4. P-MEP solves only the smallest 5 instances, finding the optimal plans. The ratio LPG-TD.speed vs HSP*a is [1.00(1.02)1.11]. The ratio 18. Note that this is not possible in the Airport domain - that can also be viewed as a type of scheduling problem (Hatzack & Nebel, 2001) - due to the restricted space on the airport. In that sense, the Airport domain incorporates more planning aspects than UMTS. 565 Hoffmann & Edelkamp 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEYPMEP LPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 TP4HSPS_A (a) (b) 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN CRIKEYLPG-3 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 TP4HSPS_A (c) (d) Figure 13: UMTS, satisficing (a) optimal (b); UMTS-flaw, satisficing (c) optimal (d). LPG-TD.quality vs HSP*a is [1.00(1.00)1.03]. The ratio LPG-3 vs HSP*a is [1.00(1.53)2.27]; we remark that the plan quality data used here is for "LPG-3.bestquality", which takes the entire available half hour time trying to optimize the found plan. Looking at the plots, one sees that the LPG-3 makespan curve is much steeper than those of the other planners. In Figure 13 (c) and (d), we see the performance of the IPC-4 planners in the same basic domain version, but with the flaw construct. LPG-TD remains unaffected, but SGPlan and CRIKEY become a lot worse. Particularly, CRIKEY now takes several minutes to solve even the smallest instances. We take this to confirm our intuition that the flaw can (but does not necessarily) confuse the heuristic functions of relaxed-plan based heuristic planners. Whether or not the heuristic function becomes confused probably depends on details in the particular way the relaxed plans are constructed (such a construction may, e.g., choose between different actions based on an estimate of how harmful they are to already selected actions). In the optimal track, when introducing the flaw construct into the domain, the performance of TP4 and HSP*a becomes slightly worse, and nearly indistinguishable. Regarding plan quality, there now is a quality difference in terms of the number of actions needed by CRIKEY and SGPlan. SGPlan's plans still all have the smallest pos566 The Deterministic Part of IPC-4: An Overview 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN TILSAPA 0.01 0.1 1 10 100 1000 10000 5 10 15 20 25 30 35 40 45 50 LPG-TDSGPLAN (a) (b) Figure 14: UMTS-timewindows (a); UMTS-flaw-timewindows (b). sible "length", bx/5c * 8. Those of CRIKEY have a lot of variance, and are longer; the ratio CRIKEY vs SGPlan is [1.07(1.72)3.88]. In the group minimizing makespan, the observations are very similar to the unflawed domain version above. The ratio LPG-TD.speed vs HSP*a is [1.00(1.01)1.04], and the ratio LPG-TD.quality vs HSP*a is [1.00(1.00)1.01]. Although it solves only the smaller half of the instances, the ratio LPG-3.bestquality vs HSP*a is [1.00(1.48)2.27]. Figure 14 shows the results in the two domain versions with explicitly encoded time windows. No optimal planners could participate since none of them could handle timed initial literals. In the version without the flaw, again LPG-TD and SGPlan need only split seconds. The only other competitor, Tilsapa, needs more runtime but also scales up well. When introducing the flaw, the only competitors are SGPlan and LPG-TD. SGPlan's runtime performance becomes a lot worse, while LPG-TD remains completely unaffected. Regarding plan quality, the number of actions in SGPlan's plans is still bx/5c*8 in all cases. LPG-TD.speed and LPG-TD.quality return plans with identical makespan in all cases. In the non-flawed version, UMTS-timewindows, the makespan ratio Tilsapa vs LPG-TD is [1.00(1.20)1.37]. We do not show runtime graphs for the domain versions with compiled time windows. In UMTS-timewindows-compiled, SGPlan and CRIKEY participated. Both scale up well and solve all instances, but SGPlan needs only split seconds, while CRIKEY needs up to more than 100 seconds per instance in the larger cases. Both planners try to minimize the number of actions, and both need exactly bx/5c * 8 + 5 actions for each instance number x - namely, the optimal number bx/5c * 8 of actions as before, plus 5 artificial actions encoding the time windows. In UMTS-flaw-timewindows-compiled, the sole participating planner was SGPlan. It solved only the smaller half of the instances, finding plans of lengthb x/5c * 8 + 6 - i.e., using one unnecessary action in each instance (the action concerns an application that does not need to be set up). In domain version UMTS, we awarded 1st places to SGPlan, LPG-TD, and HSP*a; we awarded 2nd places to CRIKEY and TP4. In domain version UMTS-flaw, we awarded 567 Hoffmann & Edelkamp a 1st place to LPG-TD, and 2nd places to SGPlan, HSP*a, and TP4. In version UMTStimewindows, we awarded 1st places to SGPlan and LPG-TD, and a 2nd place to Tilsapa. In UMTS-flaw-timewindows, we awarded a 1st place to LPG-TD. In UMTS-timewindowscompiled, we awarded a 1st place to SGPlan. 6. IPC-4 Awards The numbers of 1st and 2nd places achieved by the planners are shown in Tables 3 and 4; planners that never came in 1st or 2nd place are left out of the tables. Since many of the planners (6 of the satisficing planners, and 4 of the optimal planners) only dealt with the purely propositional domain versions (i.e., STRIPS or ADL), we counted the performance in these domains separately. SGPlan LPG-TD FD FDD Macro-FF YAHSP CRIKEY Tilsapa Propositional 3 / 6 1 / 6 5 / 2 6 / 3 3 / 0 4 / 1 0 / 0 0 / 0 Temp/Metric 13 / 2 7 / 7 0 / 1 0 / 1 Total Count 16 / 8 8 / 13 5 / 2 6 / 3 3 / 0 4 / 1 0 / 1 0 / 1 Table 3: Summary of results: satisficing planners, number of 1st places / number of 2nd places. CPT TP-4 HSP*a SATPLAN Optiplan Semsyn BFHSP Propositional 0 / 1 0 / 0 0 / 0 5 / 1 0 / 4 0 / 2 0 / 2 Temp/Metric 3 / 0 0 / 5 1 / 2 Total Count 3 / 1 0 / 5 1 / 2 5 / 1 0 / 4 0 / 2 0 / 2 Table 4: Summary of results: optimal planners, number of 1st places / number of 2nd places. For the satisficing planners, based on our observations we decided to award separate prizes for performance in the pure STRIPS and ADL domains. For the optimal planners this seemed not appropriate due to, first, the small number of planners competing in the temporal/metric domains, and, second, the smaller overall number of competing systems - giving 4 prizes to 7 systems seemed too much. Overall, the awards made in the deterministic part of IPC-4 are the following: * 1st Prize, Satisficing Propositional Track - Fast (Diagonally) Downward, by Malte Helmert and Silvia Richter * 2nd Prize, Satisficing Propositional Track - YAHSP, by Vincent Vidal * 2nd Prize, Satisficing Propositional Track - SGPlan, by Yixin Chen, Chih-Wei Hsu and Benjamin W. Wah * 1st Prize, Satisficing Metric Temporal Track - SGPlan, by Yixin Chen, Chih-Wei Hsu and Benjamin W. Wah 568 The Deterministic Part of IPC-4: An Overview * 2nd Prize, Satisficing Metric Temporal Track - LPG-TD, by Alfonso Gerevini, Alessandro Saetti, Ivan Serina, and Paolo Toninelli * 1st Prize, Optimal Track - SATPLAN, by Henry Kautz, David Roznyai, Farhad Teydaye-Saheli, Shane Neth, and Michael Lindmark * 2nd Prize, Optimal Track - CPT, by Vincent Vidal and Hector Geffner We would like to re-iterate that the awarding of prizes is, and has to be, a very sketchy "summary" of the results of a complex event such as IPC-4. A few bits of information are just not sufficient to summarize thousands of data points. Many of the decisions we took in the awarding of the prizes, i.e. in the judgement of scaling behavior, were very close. This holds especially true for most of the runtime graphs concerning optimal planners, and for some of the runtime graphs concerning satisficing propositional planners. What we think is best, and what we encourage everybody to do, is to have a closer look at the results plots for themselves. As said before, the full plots are available in an online appendix. 7. Conclusion All in all, our feeling as the organizers is that the deterministic part of IPC-4 was a great success, and made several valuable contributions to the field. To mention the, from our perspective, two most prominent points: the event provided the community with a set of interesting new benchmarks, and made visible yet another major step forward in the scalability of satisficing planning systems. The latter was made possible by novel heuristics and domain analysis techniques. There is a wide variety of questions to be addressed in the context of the future of the (deterministic part of) the IPC. Let us discuss just a few that we feel are important. Regarding benchmark domains, as said we invested significant effort into the IPC-4 benchmarks, and we would definitely recommend to re-use some of them in future IPC editions. While some domain versions were solved relatively easily by certain groups of planners, many of them still constitute major challenges. Examples are Pipesworld with tankage restrictions, Optical-Telegraph, large PSR instances, and UMTS for optimal planners. That said, most of the IPC-3 domains are also still challenging and should be re-used. It is probably more useful for the community to consolidate performance on the existing set of benchmarks, rather than to increase the benchmark database at large pace. For one thing, to measure progress between IPC editions, re-used benchmark domains (and instances), like Satellite and Settlers in the case of IPC-4, are more useful.19 More generally, the benchmark set is already large; if there is a too large set, then a situation may arise where authors select rather disjoint subsets in their individual experiments. 19. In the field addressing the SAT problem, particularly in the respective competition events, progress is also often measured simply in terms of the size (number of variables and clauses) of the formulas that could be tackled successfully, making just a few distinctions about the origin of the formulas (like, randomly generated or from an application). In the context of the IPC, one can do similar things by measuring parameters such as, e.g., the number of ground actions tackled successfully. However, given the large differences between the individual domains used in the IPC, more distinctions must be made. A detailed investigation of the effect of several parameters, in the IPC-4 domains, is given by Edelkamp et al. (2005). 569 Hoffmann & Edelkamp Regarding PDDL extensions, the formalisms for derived predicates and timed initial literals that we introduced are only first steps into the respective directions. The PDDL2.2 derived predicates formalism is restrictive in that it allows no negative interactions between derived predicates; one could easily allow negative interactions that do not lead to cycles and thus to ambiguous semantics (Thiebaux et al., 2003, 2005). One could also imagine derivation rules for values of more general data types than predicates/Booleans, particularly for numeric variables. As for timed initial literals, obviously they encode only a very restrictive subset of the wide variety of forms that exogenous events can take. Apart from exogenous events on numeric values, such events may in reality be continuous processes conditioned on the world state, rather than finitely many discrete time instants known beforehand. The PDDL2.2 action model itself is, of course, still restrictive in its postulation of discrete variable value updates. Still we believe that the IPC should not let go off the very simple PDDL subsets such as STRIPS, to support accessibility of the competition. As noted in Section 4.1, most systems are still not able to handle all the language features introduced with IPC-3 and IPC-4. Also, we believe that a STRIPS track, and more generally domain versions formulated in simple language subsets, is important to encourage basic algorithms research. A new idea is easier to try in a simple language. To avoid misunderstandings: the language features introduced with PDDL2.1 and PDDL2.2 already have a significant basis of acceptance in implemented systems, and should definitely be kept for future editions of the IPC. In the context of basic research, it should be noted that the satisficing track of IPC-4 was almost entirely populated by planners of the relaxed-plan based heuristic search type. This demonstrates the danger of the competition to concentrate research too much around a few successful methods. Still, two of the most remarkable planners in that track, FastDownward and SGPlan, use a significantly different heuristic and new domain analysis techniques as the basis of their success, respectively. Putting the optimal planners in a separate track serves to maintain this, very different, type of planning algorithms. We recommend to keep this distinction in the future IPC events. About the hand-tailored track, we have to say that it seems unclear if and how that could be brought back into the focus. Maybe a more suitable form of such an event would be an online (at the hosting conference) programming (i.e., planner tailoring) competition. This would, of course, imply a much smaller format for the event than the format the IPC for automated planners has grown to already. But an online hand-tailored competition would have the advantage of better visibility of programming efforts, and of not taking up prohibitively much time of the system developers. In the context of competition size, last not least there are a few words to be said about the role and responsibilities of the organizers. The IPC has grown too large to be handled, in all its aspects, by just two persons. It may be worth thinking about distributing the organization workload among a larger group of people. One approach that might make sense would be to let different people organize the tracks concerning the different PDDL subsets. Another approach would be to let different people handle the language definition, the benchmark preparation, and the results collection, respectively. It would probably be a good idea to establish an IPC council that, unlike the organizing committees in the past, 570 The Deterministic Part of IPC-4: An Overview would persist across individual competitions, and whose role would be to actively set up and support the organizing teams. Acknowledgments We would like to thank the IPC-4 organizing committee, namely Fahiem Bacchus, Drew McDermott, Maria Fox, Derek Long, Jussi Rintanen, David Smith, Sylvie Thiebaux, and Daniel Weld for their help in taking the decision about the language for the deterministic part of IPC-4, and in ironing out the details about syntax and semantics. We especially thank Maria Fox and Derek Long for giving us the latex sources of their PDDL2.1 article, and for discussing the modifications of this document needed to introduce the semantics of derived predicates and timed initial literals. We are indebted to Roman Englert, Frederico Liporace, Sylvie Thiebaux, and Sebastian Tr"ug, who helped in the creation of the benchmark domains. We wish to say a big thank you to all the participating teams for their efforts. There is significant bravery in the submission of a planning system to a competition, where the choice and design of the benchmark problems is up to the competition organizers, not to the individuals. We thank the LPG team for investing the extra effort of running the IPC-3 LPG version on the IPC-4 benchmarks, and we thank Shahid Jabbar for proofreading this text. We thank Subbarao Kambhampati for pointing out that the name "classical part" is ambiguous, and suggesting to use "deterministic part" instead. Last not least, we thank the anonymous reviewers, and Maria Fox in her role as the responsible JAIR editor, for their comments; they helped to improve the paper. At the time of organizing the competition, J"org Hoffmann was supported by the DFG (Deutsche Forschungsgemeinschaft), in the research project "HEUPLAN II". Stefan Edelkamp was and is supported by the DFG in the research project "Heuristic Search". Appendix A. BNF Description of PDDL2.2 This appendix contains a complete BNF specification of the PDDL2.2 language. For readability, we mark with (* * *) the points in the BNF where, in comparison to PDDL2.1, the new language constructs of PDDL2.2 are inserted. A.1 Domains Domains are defined exactly as in PDDL2.2, except that we now also allow to define rules for derived predicates at the points where operators (actions) are allowed. hdomaini ::= (define (domain hnamei)[hrequire-defi] [htypes-defi]:typing[hconstants-defi] [hpredicates-defi][hfunctions-defi] :fluentsh structure-defi*)hrequire-defi ::= (:requirements hrequire-keyi +)h require-keyi ::= See Section A.6 571 Hoffmann & Edelkamp htypes-defi ::= (:types htyped list (name)i)hconstants-defi ::= (:constants htyped list (name)i) hpredicates-defi ::= (:predicates hatomic formula skeletoni+)hatomic formula skeletoni ::= (hpredicatei htyped list (variable)i)hpredicatei ::= hnamei hvariablei ::= ?hnameihatomic function skeletoni ::= (hfunction-symboli htyped list (variable)i)hfunction-symboli ::= hnamei hfunctions-defi ::=:fluents (:functions hfunction typed list(atomic function skeleton)i) hstructure-defi ::= haction-defihstructure-defi ::= :durative-actions hdurative-action-defi (* * *) hstructure-defi ::=:derived-predicates hderived-defihtyped list ( x)i ::= x*htyped list ( x)i ::=:typing x+- htypei htyped list(x)ihprimitive-typei::= hnamei htypei ::= (either hprimitive-typei+)htypei ::= hprimitive-typei hfunction typed list (x)i ::= x*hfunction typed list ( x)i ::=:typing x+- hfunction typeihfunction typed list( x)ihfunction typei ::= number A.2 Actions The BNF for an action definition is the same as in PDDL2.2. haction-defi ::= (:action haction-symboli:parameters ( htyped list (variable)i ) haction-def bodyi)haction-symboli ::= hnamei haction-def bodyi ::= [:precondition hGDi][:effect heffecti] hGDi ::= ()hGDi ::= hatomic formula(term)i hGDi ::=:negative-preconditions hliteral(term)ihGDi ::= (and hGDi*) hGDi ::=:disjunctive-preconditions (or hGDi*)hGDi ::= :disjunctive-preconditions (not hGDi)h GDi ::=:disjunctive-preconditions (imply hGDi hGDi)hGDi ::= :existential-preconditions (exists (htyped list(variable)i*) hGDi )hGDi ::= :universal-preconditions (forall (htyped list(variable)i*) hGDi )hGDi ::= :fluents hf-compih f-compi ::= (hbinary-compi hf-expi hf-expi)hliteral( t)i ::= hatomic formula(t)ihliteral( t)i ::= (not hatomic formula(t)i) 572 The Deterministic Part of IPC-4: An Overview hatomic formula(t)i ::= (hpredicatei t*)htermi ::= hnamei htermi ::= hvariableihf-expi ::= hnumberi hf-expi ::= (hbinary-opi hf-expi hf-expi)hf-expi ::= (- hf-expi) hf-expi ::= hf-headihf-headi ::= (hfunction-symboli htermi*) hf-headi ::= hfunction-symbolihbinary-opi ::= + hbinary-opi ::= -hbinary-opi ::= * hbinary-opi ::= /hbinary-compi ::= >hbinary-compi ::==hbinary-compi ::= <=hnumberi ::= Any numeric literal (integers and floats of form n.n).heffecti ::= () heffecti ::= (and hc-effecti*)heffecti ::= hc-effecti hc-effecti ::=:conditional-effects (forall (hvariablei*) heffecti)hc-effecti ::= :conditional-effects (when hGDi hcond-effecti)h c-effecti ::= hp-effectihp-effecti ::= (hassign-opi hf-headi hf-expi) hp-effecti ::= (not hatomic formula(term)i)hp-effecti ::= hatomic formula(term)i hp-effecti ::=:fluents(hassign-opi hf-headi hf-expi)hcond-effecti ::= (and hp-effecti*) hcond-effecti ::= hp-effectihassign-opi ::= assign hassign-opi ::= scale-uphassign-opi ::= scale-down hassign-opi ::= increasehassign-opi ::= decrease A.3 Durative Actions Durative actions are the same as in PDDL2.2, except that we restrict ourselves to level 3 actions, where the duration is given as the fixed value of a numeric expression (rather than as the possible values defined by a set of constraints). This slightly simplifies the BNF. hdurative-action-defi ::= (:durative-action hda-symboli:parameters ( htyped list (variable)i ) hda-def bodyi)hda-symboli ::= hnamei hda-def bodyi ::= :duration (= ?duration hf-expi):condition hda-GDi 573 Hoffmann & Edelkamp :effect hda-effectihda-GDi ::= () hda-GDi ::= htimed-GDihda-GDi ::= (and htimed-GDi +)h timed-GDi ::= (at htime-specifieri hGDi)htimed-GDi ::= (over hintervali hGDi) htime-specifieri ::= starthtime-specifieri ::= end hintervali ::= all A.4 Derived predicates As said, rules for derived predicates can be given in the domain description at the points where actions are allowed. The BNF is: (* * *) hderived-defi ::= (:derived htyped list (variable)i hGDi) Note that we allow the specification of types with the derived predicate arguments. This might seem redundant as the predicate types are already declared in the :predicates field. Allowing to specify types with the predicate (rule) "parameters" serves to give the language a more unified look-and-feel, and one might use the option to make the parameter ranges more restrictive. (Remember that the specification of types is optional, not mandatory.) Repeating what has been said in Section 3.1.1, this BNF is more generous than what is considered a well-formed domain description in PDDL2.2. We call a predicate P derived if there is a rule that has a predicate P in its head; otherwise we call P basic. The restrictions we apply are: 1. The actions available to the planner do not affect the derived predicates: no derived predicate occurs on any of the effect lists of the domain actions. 2. If a rule defines that P (x) can be derived from OE(x), then the variables in x are pairwise different (and, as the notation suggests, the free variables of OE(x) are exactly the variables in x). 3. If a rule defines that P (x) can be derived from OE, then the Negation Normal Form (NNF) of OE(x) does not contain any derived predicates in negated form. A.5 Problems The only change made to PDDL2.1 in the problem description is that we allow the specification of timed initial literals. hproblemi ::= (define (problem hnamei)(:domain hnamei) [hrequire-defi][hobject declarationi ] hiniti 574 The Deterministic Part of IPC-4: An Overview hgoali[hmetric-speci] [hlength-speci ])hobject declarationi ::= (:objects htyped list (name)i) hiniti ::= (:init hinit-eli*)hinit-eli ::= hliteral(name)i hinit-eli ::=:fluents (= hf-headi hnumberi)(* * *) hinit-eli ::= :timed-initial-literals (at hnumberi hliteral(name)i)h goali ::= (:goal hGDi)hmetric-speci ::= (:metric hoptimizationi hground-f-expi) hoptimizationi ::= minimizehoptimizationi ::= maximize hground-f-expi ::= (hbinary-opi hground-f-expi hground-f-expi)hground-f-expi ::= (- hground-f-expi) hground-f-expi ::= hnumberihground-f-expi ::= (hfunction-symboli hnamei*) hground-f-expi ::= total-timehground-f-expi ::= hfunction-symboli Repeating what has been said in Section 3.1.1, the requirement flag for timed initial literals implies the requirement flag for durational actions (see also Section A.6), i.e. the language construct is only available in PDDL2.2 level 3. Also, the above BNF is more generous than what is considered a well-formed problem description in PDDL2.2. The times hnumberi at which the timed literals occur are restricted to be greater than 0. If there are also derived predicates in the domain, then the timed literals are restricted to not influence any of these, i.e., like action effects they are only allowed to affect the truth values of the basic (non-derived) predicates (IPC-4 will not use both derived predicates and timed initial literals within the same domain). A.6 Requirements Here is a table of all requirements in PDDL2.2. Some requirements imply others; some are abbreviations for common sets of requirements. If a domain stipulates no requirements, it is assumed to declare a requirement for :strips. 575 Hoffmann & Edelkamp Requirement Description :strips Basic STRIPS-style adds and deletes :typing Allow type names in declarations of variables :negative-preconditions Allow not in goal descriptions :disjunctive-preconditions Allow or in goal descriptions :equality Support = as built-in predicate :existential-preconditions Allow exists in goal descriptions :universal-preconditions Allow forall in goal descriptions :quantified-preconditions = :existential-preconditions+ :universal-preconditions :conditional-effects Allow when in action effects :fluents Allow function definitions and use of effects usingassignment operators and arithmetic preconditions. :adl = :strips + :typing+ :negative-preconditions+ :disjunctive-preconditions+ :equality+ :quantified-preconditions+ :conditional-effects :durative-actions Allows durative actions.Note that this does not imply :fluents. :derived-predicates Allows predicates whose truth value isdefined by a formula :timed-initial-literals Allows the initial state to specify literalsthat will become true at a specified time point implies durative actions (i.e. applicable onlyin PDDL2.2 level 3) References Bacchus, F. (2000). Subset of PDDL for the AIPS2000 Planning Competition. The AIPS-00 Planning Competition Committee. Bacchus, F. (2001). The AIPS'00 planning competition. The AI Magazine, 22 (3), 47-56. Bertoli, P., Cimatti, A., Slaney, J., & Thi'ebaux, S. (2002). Solving power supply restoration problems with planning via symbolic model-checking. In Harmelen, F. V. (Ed.), Proceedings of the 15th European Conference on Artificial Intelligence (ECAI-02), pp. 576-580, Lyon, France. Wiley. Blum, A. L., & Furst, M. L. (1995). Fast planning through planning graph analysis. In Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI95), pp. 1636-1642, Montreal, Canada. Morgan Kaufmann. Blum, A. L., & Furst, M. L. (1997). Fast planning through planning graph analysis. Artificial Intelligence, 90 (1-2), 279-298. Bonet, B., & Thiebaux, S. (2003). GPT meets PSR. In Proceedings of the 13th International Conference on Automated Planning and Scheduling (ICAPS-03), pp. 102-111, Trento, Italy. Morgan Kaufmann. Cesta, A., & Borrajo, D. (Eds.). (2001). Recent Advances in AI Planning. 6th European Conference on Planning (ECP'01), Toledo, Spain. Springer-Verlag. 576 The Deterministic Part of IPC-4: An Overview Clarke, E. M., Grumberg, O., & Peled, D. (2000). Model Checking. MIT Press. Edelkamp, S. (2003a). Limits and possibilities of PDDL for model checking software.. In Edelkamp, & Hoffmann (Edelkamp & Hoffmann, 2003). Edelkamp, S. (2003b). Promela planning. In Proceedings of Model Checking Software (SPIN), pp. 197-212. Edelkamp, S. (2003c). Taming numbers and durations in the model checking integrated planning system. Journal of Artificial Intelligence Research, 20, 195-238. Edelkamp, S., & Hoffmann, J. (Eds.). (2003). Proceedings of Workshop on the Competition: Impact, Organization, Evaluation, Benchmarks, at ICAPS'03. AAAI Press. Edelkamp, S., Hoffmann, J., Englert, R., Liporace, F., Thiebaux, S., & Tr"ug, S. (2005). Engineering benchmarks for planning: the domains used in the deterministic part of IPC-4. Journal of Artificial Intelligence Research. Submitted. Edelkamp, S., Hoffmann, J., Littman, M., & Younes, H. (Eds.). (2004). Proceedings of the 4th International Planning Competition. JPL. Englert, R. (2003). Re-scheduling with temporal and operational resources for the mobile execution of dynamic UMTS applications. In KI-Workshop AI in Planning, Scheduling, Configuration and Design (PUK). Englert, R. (2005). Planning to optimize the UMTS call set-up for the execution of mobile agents. Journal of Applied Artificial Intelligence (AAI), 19 (2), 99-117. Fikes, R. E., & Nilsson, N. (1971). STRIPS: A new approach to the application of theorem proving to problem solving. Artificial Intelligence, 2, 189-208. Fox, M., Long, D., & Halsey, K. (2004). An investigation into the expressive power of PDDL2.1. In Saitta, L. (Ed.), Proceedings of the 16th European Conference on Artificial Intelligence (ECAI-04), Valencia, Spain. Wiley. Fox, M., & Long, D. (2003). PDDL2.1: An extension to PDDL for expressing temporal planning domains. Journal of Artificial Intelligence Research, 20, 61-124. Gerevini, A., Saetti, A., & Serina, I. (2003). Planning through stochastic local search and temporal action graphs. Journal of Artificial Intelligence Research, 20, 239-290. Haslum, P., & Geffner, H. (2001). Heuristic planning with time and resources.. In Cesta, & Borrajo (Cesta & Borrajo, 2001), pp. 121-132. Hatzack, W., & Nebel, B. (2001). The operational traffic control problem: Computational complexity and solutions.. In Cesta, & Borrajo (Cesta & Borrajo, 2001), pp. 49-60. Helmert, M. (2001). On the complexity of planning in transportation domains.. In Cesta, & Borrajo (Cesta & Borrajo, 2001), pp. 349-360. Helmert, M. (2003). Complexity results for standard benchmark domains in planning. Artificial Intelligence, 143, 219-262. Helmert, M. (2004). A planning heuristic based on causal graph analysis. In Koenig, S., Zilberstein, S., & Koehler, J. (Eds.), Proceedings of the 14th International Conference on Automated Planning and Scheduling (ICAPS-04), pp. 161-170, Whistler, Canada. Morgan Kaufmann. 577 Hoffmann & Edelkamp Hipke, C. A. (2000). Distributed Visualization of Geometric Algorithms. Ph.D. thesis, University of Freiburg. Hoffmann, J. (2003). The Metric-FF planning system: Translating "ignoring delete lists" to numeric state variables. Journal of Artificial Intelligence Research, 20, 291-341. Hoffmann, J. (2005). Where ignoring delete lists works: Local search topology in planning benchmarks. Journal of Artificial Intelligence Research. To appear. Hoffmann, J., & Nebel, B. (2001). The FF planning system: Fast plan generation through heuristic search. Journal of Artificial Intelligence Research, 14, 253-302. Holzmann, G. J. (2004). The SPIN model checker: Primer and reference manual. Addison Wesley. Long, D., & Fox, M. (2003). The 3rd international planning competition: Results and analysis. Journal of Artificial Intelligence Research, 20, 1-59. McDermott, D. (2000). The 1998 AI planning systems competition. The AI Magazine, 21 (2), 35-55. McDermott, D., et al. (1998). The PDDL Planning Domain Definition Language. The AIPS-98 Planning Competition Committee. Milidi'u, R. L., & dos Santos Liporace, F. (2004). Plumber, a pipeline transportation planner. In Proceedings of the International Workshop on Harbour and Maritime Simulation, HMS, pp. 99-106, Rio de Janeiro, Brazil. Milidiu, R. L., dos Santos Liporace, F., & de Lucena, C. J. (2003). Pipesworld: Planning pipeline transportation of petroleum derivatives.. In Edelkamp, & Hoffmann (Edelkamp & Hoffmann, 2003). Pednault, E. P. (1989). ADL: Exploring the middle ground between STRIPS and the situation calculus. In Brachman, R., Levesque, H. J., & Reiter, R. (Eds.), Principles of Knowledge Representation and Reasoning: Proceedings of the 1st International Conference (KR-89), pp. 324-331, Toronto, ON. Morgan Kaufmann. Penberthy, J. S., & Weld, D. S. (1992). UCPOP: A sound, complete, partial order planner for ADL. In Nebel, B., Swartout, W., & Rich, C. (Eds.), Principles of Knowledge Representation and Reasoning: Proceedings of the 3rd International Conference (KR92), pp. 103-114, Cambridge, MA. Morgan Kaufmann. Thi'ebaux, S., & Cordier, M.-O. (2001). Supply restoration in power distribution systems -- a benchmark for planning under uncertainty.. In Cesta, & Borrajo (Cesta & Borrajo, 2001), pp. 85-95. Thi'ebaux, S., Cordier, M.-O., Jehl, O., & Krivine, J.-P. (1996). Supply restoration in power distribution systems -- a case study in integrating model-based diagnosis and repair planning. In Horvitz, E., & Jensen, F. (Eds.), Proceedings of the 12th International Conference on Uncertainty in AI (UAI-96), pp. 525-532, Portland, OR, USA. Morgan Kaufmann. Thiebaux, S., Hoffmann, J., & Nebel, B. (2003). In defence of PDDL axioms. In Gottlob, G. (Ed.), Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI-03), pp. 961-966, Acapulco, Mexico. Morgan Kaufmann. 578 The Deterministic Part of IPC-4: An Overview Thiebaux, S., Hoffmann, J., & Nebel, B. (2005). In defence of PDDL axioms. Artificial Intelligence. To appear. Tr"ug, S., Hoffmann, J., & Nebel, B. (2004). Applying automatic planning systems to airport ground-traffic control -- a feasibility study. In Biundo, S., Fr"uhwirth, T., & Palm, G. (Eds.), KI-04: Advances in Artificial Intelligence, pp. 183-197, Ulm, Germany. Springer-Verlag. Vidal, V., & Geffner, H. (2004). Branching and pruning: An optimal temporal POCL planner based on constraint programming. In Proceedings of the 19th National Conference of the American Association for Artificial Intelligence (AAAI-04), pp. 570-577, San Jose, CA. MIT Press. Younes, H., & Littman, M. (2005). Probabilistic part of the 4th international planning competition. Journal of Artificial Intelligence Research. Submitted. 579