Automatic Creation of Taxonomies of Genetic Programming Systems

Abstract

Abstract. A few attempts to create taxonomies in evolutionary computation have been made. These either group algorithms or group problems on the basis of their similarities. Similarity is typically evaluated by manually analysing algorithms/problems to identify key characteristics that are then used as a basis to form the groups of a taxonomy. This task is not only very tedious but it is also rather subjective. As a consequence the resulting taxonomies lack universality and are sometimes even questionable. In this paper we present a new and powerful approach to the construction of taxonomies and we apply it to Genetic Programming (GP). Only one manually constructed taxonomy of problems has been proposed in GP before, while no GP algorithm taxonomy has ever been suggested. Our approach is entirely automated and objective. We apply it to the problem of grouping GP systems with their associated parameter settings. We do this on the basis of performance signatures which represent the behaviour of each system across a class of problems. These signatures are obtained thorough a process which involves the instantiation of models of GP's performance. We test the method on a large class of Boolean induction problems.

Theoretical results in genetic programming: the next ten years?

“…That is, they cannot easily be used to estimate, for example, the success rate of a GP system or the expected number of fitness evaluations required to solve a problem. Recent work [86,87] has begun to rectify this situation. The work has found that a simple semilinear model can accurately predict the performance of program-induction algorithms for the domains of symbolic regression and Boolean function induction.…”

“…2.5, the simple performance modelling techniques for GP proposed in [86,87] offer an alternative way of assessing performance, and, hence, difficulty for GP. While these models are simple and approximate, they have provided accurate predictions.…”

“…As we mentioned in Sect. 2.5, it is possible to use simple models of GP [86,87] to make accurate predictions of performance and use them to construct useful taxonomies of evolutionary program-induction algorithms for specific classes of problems. A technique to produce taxonomies of problems in evolutionary computation (including GP) has also been proposed in [132,133].…”