close this bookVolume 6: No. 84
View the documentFunding news
View the documentNew architectures
View the documentIndustry news
View the documentInput/translation technologies
View the documentResearch software (in our CRS 6.42 digest this week)
View the documentFuzzy systems
View the documentGenetic algorithms
View the documentNeural networks

DOE's Accelerated Strategic Computing Initiative (ASCI) consolidates the modeling and simulation efforts of LLNL, LANL, and Sandia National Laboratory, particularly with regard to nuclear stockpile management. The labs form alliances and Centers of Excellence with universities, for work on key multidisciplinary research. One of the goals is to "accelerate advances in critical basic sciences, mathematics and computer science areas, in computational science and engineering, in high performance computing systems and in problem solving environments" supporting long-term ASCI objectives. This is done through academic strategic alliance centers (over 5-10 years), strategic investigations, and individual collaborations, with all work unclassified and publishable. For instance, a center might address scientific data management, visualization, or software tools for terascale computing environments. At least five such centers are expected to be started in FY 97 and FY 98, funded for $1.5M to $2M per center for the first year (growing to $4M-$5M/year). Strategic investigations will be funded at $100K-$400K/year, for a total of $3M in FY 97. Individual collaborations are to be funded at $50K-$100K out of laboratory ASCI programmatic funds. .

LLNL has issued an ASCI Centers Request For Expression of Interest and Preliminary Proposals, due 1/9/97. Preliminary proposals should focus on how existing research efforts integrate with implementation of an Alliance Center of Excellence. Final proposals are due 3/18/97. Technical queries may be sent to Dona Crawford , Dick Watson , or Ann Hayes ; business and proposal queries to Lynn Rippe , (510) 423-2176. [Merrell Patrick , 11/25/96.]

A GAO report found research costs at public colleges and universities up 157% from 1980 to 1994, the fastest-growing contributor to a 234% tuition increase. State appropriations were declining during this time, while household income rose only 82%. Indirect cost recovery on grants has been capped at 26% since 1991, further driving up tuition. [Robert L. Park, WHAT'S NEW, 11/22/96.] (I'm not convinced that government funding of research and infrastructure is superior to industry investment, or to "venture funding" or tenure perks by the universities themselves. It's OK for the government to grease the skids where profit motive is lacking, or where the government itself needs a service, but easy overhead money only encourages growth of overhead. The cost of research goes up because we expect it to, more often than because there is no way to do research inexpensively. The same thing happens to business costs or to lifestyle costs if money seems easily available.)

Two UTokyo researchers have built a supercomputer for simulating star cluster dynamics with up to 32K bodies. Their "gravity pipe" uses 1,692 copies of a custom chip, at a cost of just $1.5M over two years. [Popular Science, 12/96, p. 32. EDUPAGE.]

For a somewhat bizarre analog processor, see Jonathan Mills' "The Continuous Retina: Image Processing with a Single-Sensor Artificial Neural Field Network" in Proc. 1996 Int. Conf. on Neural Networks. Mill's chip solves the diffusion equation (Laplacian) using charge gradient in a conductive sheet (n-well inside a diode guard ring), sampled to generate a current input to piecewise-linear functions stored as digitally reconfigurable continuous-valued logic functions (similar to fuzzy logic). Only the latter functions employ transistors, on about 1/3 of the chip. The chip can be used as a silicon retina or for any physical computation modeled by diffusion -- including certain neural networks (without neurons!). Although not efficient for all problems, the programmable processors are universal in a theoretical sense. [, comp.arch, 11/27/96.]

("Inside every digital circuit, there's an analog signal screaming to get out." -- Al Kovalick, Hewlett-Packard.)

Moshe Sipper has a new page on "Cellular Programming: Evolution of Parallel Cellular Machines." . [, comp.theory.cell-automata, 11/22/96.]

Science Applications International Corp. (SAIC) is acquiring Bellcore, for about $700M. [WSJ, 11/22/96, B6. EDUPAGE.]

Columnist Don Crabb quotes a venture capitalist saying that Windows CE (WinCE) devices will fail to attract buyers. People don't want small general-purpose computers. "As Apple has proven with the Newton OS and U.S. Robotics with their Pilot PDAs, you have to have these devices very clearly defined and articulated. And porting Windows to them, is neither. In a year, Windows CE will be yet another dead PDA format." [NewtNews, 11/26/96. Bill Park.] (Wired magazine apparently agrees. I heard that they've moved WinCE to their "Not Hot" category.)

Mac clone maker Power Computing has licensed BeOS from Jean-Louis Gassee's Be, Inc., to ship with MacOS on every computer (exclusive until 4/97). BeOS has previously been available only at . Switching between MacOS and BeOS will require a full boot, possibly taking several minutes. (A MacOS emulator is needed, and would allow cutting and pasting between processes in different OS windows.) BeOS is said to be fast and robust -- a worthy competitor for Windows NT -- and has multithreading, preemptive multiprocessing, and an integrated object-oriented OS/application architecture. There's almost no software for it yet, and this give-away by Power Computing will help break the chicken/egg problem. (It also puts pressure on Apple to acquire BeOS or to come out with an alternative. If Apple doesn't buy Be, one of several competitors might.) Over 1K developers are working on number-crunching applications in graphic design, rendering, animation, music synthesis, speech and image processing, video editing, 3D virtual worlds, and high-bandwidth network servers. Gassee' was pleased by a comment that his Be Box was "just a poor man's Silicon Graphics" after Be licensed SGI's OpenGL industry-standard 3D application programming interface (API), to be integrated into BeOS. [Bill Park , based on Reuters and, 11/26/96.]

Ike Nassi, the head of Apple's OS division, has resigned to pursue other interests. [NewtNews, 11/26/96. Bill Park.] (It's generally agreed that Copland (intended as MacOS 8) is dead, with pieces to be released as upgrades for MacOS 7. Apple's next hope is Gershwin, to incorporate many of the same features as BeOS. It's still a long way from release.)

Handwriter Manta is a pen input device for Windows. $250 from Communication Intelligence Corp., , (888) 832-0242. [Mobile Computing and Communications, 11/96. NewtNews.]

The pen computer market is now $1B/year. . [Pen Computing, 9/96. NewtNews.]

Acer plans to license the "WisdomPen" Chinese Handwriting Recognition System from Motorola's Lexicus Division, for the Taiwanese market. Lexicus has also announced software able to recognize continuous Chinese speech. [iNews. NewtNews, 11/12/96 and 11/19/96.]

Apple has released their Mac OS Cantonese Dictation Kit 1.0. . [iNews. NewtNews, 11/12/96.]

Voice Pilot is a speech recognition system able to transcribe dictation (e.g., online chat) at 140 wpm, and to translate to other languages. Available mid-1997 on OS/2, then Windows, for $200-$300. Developed by Voice Pilot Technologies and IBM. Rolph Rudestam, 909-585-6122. [Patrick McKenna,, 11/21/96. Bill Park.]

Kamejima Co. Ltd. (Tokyo) now offers E-J Bank and J-E Bank Japanese translation software, for 50K yen. The Windows 95/NT packages have a 500K-word dictionary and 7K grammar rules. , +81-3-3798-4838, +81-3-3798-4839 Fax. [Martyn Williams,, 11/21/96. Bill Park.]

Gamelan Java applets for GA, GP, CA, TSP, SOM, NN, and ML.

SSS and AGABIL: experimental GA/ML C code.

AISearch II: C++ library of AI search algorithms.

Net++ 1.1: NN simulator for Windows.

ainet122: NN DLL development kit for Win 95.

RIPPER: noise-tolerant rule-learning system, adapted for text classification.

dtSearch v4.1: full-text search with stemming, phonetics, proximity, etc.

Data Natural (dn60): NL data storage/retrieval based on cognitive simulation.

PROGRES: graph-rewriting visual DB programming language.

MAULE Machine 1.0: extensible JavaScript programmable calculator.

Slice5: data visualization tool.

TerrainMap 2.0: 3D surface models from points, grids, or contours.

Windmill: data acquisition and control software demos.

Thinking Logs Module 1 - Lite (TLM1-Lite): NN demo for petrophysics.

CLP(FD) VLSI multilayer channel router (benchmark).

Berkeley Logo 3.4: bug-fix upgrade.

4i2i FEC: C++ library for Reed-Solomon error-correcting block encoding.

PERENTIE: ActiveX Web database server.

A famous result by Daniel Elkan showed that the two equivalent logical expressions NOT (A AND NOT B) and B OR (NOT A AND NOT B) give differing results under Zadeh max-min fuzzy logic. The same is true of Lukasiewicz logic and probabilistic logic. Elkan's contradiction is resolved in a paper by James J. Buckley and William Siler, soon to be available. The key is to define A AND B = A*B + r*sa*sb and A OR B = A + B - A*B - r*sa*sb, where r between -1 and +1 represents prior association between A and B, sa = sqrt(A(1-A)), and sb = sqrt(B(1-B)). [,, 10/25/96.]

Inform Software Corp. is offering a fuzzy logic Internet server for their fuzzyTECH development system demos, benchmarks, and literature, and for links to other fuzzy-logic and neurofuzzy design resources (e.g., Intel's fuzzy application book, Motorola's new fuzzy-instruction enhanced microcontrollers, and Wonderware's new InTouch FuzzyModule). . [Christiane Melcher ,, 11/22/96.]

Aptronix, Inc. has a demo of fuzzy logic in Java at . [,, 11/23/96.]

There's a FAQ called "The Hitch-Hiker's Guide to Evolutionary Computation," at /news.answers/ai-faq/genetic/. For genetic algorithms in GIS, check out or Tim Duncan's page on vehicle routing, at . A good source on genetic programming is Jaime Fernandez's "Genetic Programming Notebook" at>. [, comp.infosystems.gis, 11/19/96.]

Nova Genetica has added 300 new links to genetic algorithm resources. A search utility will soon be implemented. . [Darin R. Molnar ,, 11/21/96.]

Thomas G. Dietterich's "Statistical Tests for Comparing Supervised Classification Learning Algorithms" compares five tests of learning algorithms. He concludes that McNemar's test is best for algorithms executed only once (e.g., because of time or cost constraints); a new 5x2cv test of 5 iterations of 2-fold cross-validation is slightly more powerful for repeated trainings. . [, connectionists, 10/16/96.]

Don't forget the FAQ, with seven sections covering Intro, Learning, Generalization, Books (& data), Free Software, Commercial Software, and Hardware. , or browse /pub/usenet/news.answers/ai-faq/neural-nets on [Warren Sarle ,, 11/22/96.]

Bob Massey offers an interactive NN demo based on maximizing entropy. . [, comp.theory, 11/22/96.]

Here's a useful application: NeuroDiet 1.0 asks you to enter the foods you eat each day and any medical symptoms you experience. It trains a neural network and can then advise you on how your health depends on the foods you eat. (Are you getting headaches from eating peanuts? Red wine? Aged cheese? Something like this could also track weather factors and pollen-related allergies.) NeuroDiet is Windows 95 shareware from Stephen Wolstenholme , on or . [, 10/9/96.]

-- Ken