Artificial Intelligence Research Group

Some Past and Current Projects
Under Development

10. Evolutionary Algorithms for Machine Learning
Investigator(s): Dr Simon Thompson and Professor Max Bramer

Evolutionary algorithms are naturally parallel search systems that have attracted increasing levels of interest in recent years as improvements in hardware technologies has made there use more attractive. An evolutionary algorithm is any algorithm that uses a population of solutions as the basis of an iterative search, popular variants of these include Genetic Algorithms, Evolutionary Programming, Genetic Programming Classifier Systems and Evolution Strategies. We are investigating mechanisms for using these algorithms in a variety of learning tasks, in an attempt to identify their weaknesses and strengths in comparison to traditional inductive algorithms. Of particular interest is exploiting the Evolutionary Algorithms inherent parallelism to improve the performance of learning algorithms. Two main paradigms are under investigation: massively parallel genetic algorithms for solving subsymolic tasks using SIMD parallel computers; and agent based learning algorithms for dealing with very large physically distributed noisy datasets, similar to those found in corporate databases. MPGAIA (Massively Parallel Genetic Algorithm for Image Analysis) has been implemented on the departments DAP 510 parallel computer. This system aims to utilise the power of a SIMD style computer to learn image filters which can be used to discover cracks in images of aircraft fuselages. ABEL (Agent Based Evolutionary Learner) is being implemented on a network of Unix workstations. Abel is designed to provide mechanisms for dealing with so called mega-induction tasks through novel methods of problem decomposition. Issues of data fusion and data distribution are also addressed with this approach.

Thompson, S.G. Bramer, M.A. and Kalus, A. (1996) "MPGAIA- A Massively Parallel Genetic Algorithm for Image Analysis" Working Notes, AISB Workshop on Evolutionary Computing, April1-2 Brighton, Sussex. PP268-278