Setup of a New Adaptive Fuzzy Particle Swarm Optimization Algorithm

Abstract

Heuristic optimization methods such as Particle Swarm Optimization (PSO) depend on their parameters to achieve good performance on a given class of problems. Some modifications of heuristic algorithms aim to adapt those parameters during the optimization process. We present a framework to design such adaptation strategies using continuous fuzzy feedback control. Our framework, which is not tied to a particular algorithm, provides us with a simple interface where probes are sampled in the optimization process and parameters are fed back. The process of turning probes into parameters uses fuzzy logic rule sets, where the design of rules aims to maximize performance on a training benchmark. This meta-optimization is achieved by a Bayesian Optimizer (BO) with a Gradient Boosted Regression Trees (GBRT) prior. The robustness of the control is also assessed on a validation benchmark.