Pareto-coevolutionary genetic programming for problem decomposition in multi-class classification

Abstract

A bid-based approach for coevolving Genetic Programming classifiers is presented. The approach coevolves a population of learners that decompose the instance space by way of their aggregate bidding behaviour. To reduce computation overhead, a small, relevant, subset of training exemplars is (competitively) coevolved alongside the learners. The approach solves multi-class problems using a single population and is evaluated on three large datasets. It is found to be competitive, especially compared to classifier systems, while significantly reducing the computation overhead associated with training.

Managing team-based problem solving with symbiotic bid-based genetic programming

“…In contrast to traditional voting or weighted combination schemes, bid-based mechanisms have been proposed to coordinate the actions of team members designed to fulfill orthogonal roles [12,13]. Since only the 'winning' individual acts at any one time, the problem faced by ensemble methods of combining the outputs of team members is a nonissue.…”

“…Since only the 'winning' individual acts at any one time, the problem faced by ensemble methods of combining the outputs of team members is a nonissue. These approaches use GP to evolve the bidding behaviour and not to act as the classification model itself and can therefore be naturally applied to multi-class (as opposed to binary) classification problems [13].…”

“…These operators were applied with independently specified probabilities, and in all cases, a uniform distribution was used in selecting program elements. This scheme follows the GP implementation of [13].…”

“…In this work we build on the bid-based model of coevolution [12,13] which re-casts the role of GP individuals by using them to evolve a bidding strategy. The scalar actions in this model are selected from a possibly infinite set and assigned at initialization, making the approach most appropriate for domains with discrete action spaces such as classification and reinforcement learning.…”

“…The scalar actions in this model are selected from a possibly infinite set and assigned at initialization, making the approach most appropriate for domains with discrete action spaces such as classification and reinforcement learning. Earlier examples of bid-based GP considered the influence of different auction models for credit assignment [12] and the effect of using competitive coevolution to scale the approach to large datasets or problem spaces [13]. In both cases, in order to establish the 'winner' on any bidding round all members of the population had to be evaluated.…”