page 1  (6 pages)
2to next section

EPCC Seminars, Spring 1992

Felicity A. W. George
Edinburgh Parallel Computing Centre

Thursday 9th January

Pablo Moscato (University of La Plata, Argentina)
Population Approaches for Optimization, Genetic and Memetic Algorithms: The role of Hierarchical Cost Functions

Population approaches suitable for global combinatorial optimization are discussed. They are composed of a number of distinguishable individuals called agents", each one using a particular optimization strategy. Periods of independent search follow phases on which the population is restarted from new configurations. Due to its intrinsic parallelism and the asynchronicity of the method, it would be applicable for parallel computers. Results on two test problems, the Travelling Salesman Problem and issues on learning on Binary Perceptrons will be presented. The latter can be viewed as the simplest toy model for problems that arise in learning in artificial neural networks. The relevance of the Memetic approach instead of the classical Genetic algorithm will also be discussed as well as the possible application of these ideas into other fields of optimization.

Thursday 16th January

Arne ter Laak (Computer Science)
Nonconvex Continuous Optimization Experiments on a Transputer System

In this talk a systolic parallel implementation of the Fast Simulated Annealing algorithm is discussed. The continuous optimization problem we want to tackle results from large scale interactions of simple charges on a sphere. This problem model is a simplification of sperical crystallization of N (102 < N < 106) unbounded molecules. Efficiency of the implementation is improved by introducing geometrical decomposition of the costfunction calculation, i.e. determining the potential energy of the N particles, Results of the hybrid approach, algorithmic decomposition of Fast SA combined with geometrical decomposition, obtained with implementations on the Meiko are discussed.

James Mills (EPCC)
A Common Highlevel Interface to Message Passing

Parallel computing has suffered greatly from a lack of standards. In particular, each manufacturer has traditionally supplied its own message-passing system for distributed MIMD machines. This locks the customer in with the particular com-