A Feature-Based Prediction Model of Algorithm Selection for Constrained Continuous Optimisation

TL;DR

This paper develops a feature-based prediction model for selecting the most effective bio-inspired algorithm in constrained continuous optimization, demonstrating that evolved and random instances improve prediction accuracy.

Contribution

It introduces a novel approach combining evolved and random problem instances to enhance algorithm selection models in constrained optimization.

Findings

Multi-objective evolved instances improve prediction accuracy.

Combining evolved and random instances yields robust models.

Models accurately predict best algorithms across diverse problems.

Abstract

With this paper, we contribute to the growing research area of feature-based analysis of bio-inspired computing. In this research area, problem instances are classified according to different features of the underlying problem in terms of their difficulty of being solved by a particular algorithm. We investigate the impact of different sets of evolved instances for building prediction models in the area of algorithm selection. Building on the work of Poursoltan and Neumann [11,10], we consider how evolved instances can be used to predict the best performing algorithm for constrained continuous optimisation from a set of bio-inspired computing methods, namely high performing variants of differential evolution, particle swarm optimization, and evolution strategies. Our experimental results show that instances evolved with a multi-objective approach in combination with random instances of the underlying problem allow to build a model that accurately predicts the best performing algorithm for a wide range of problem instances.