Analysis of learning-based multi-agent simulated annealing algorithm for function optimisation problems

Abstract

Canonical simulated annealing SA algorithm is extremely slow in convergence, and the implementation and efficiency of parallel SA algorithms are typically problem-dependent. Multi-agent SA MSA algorithms, which use learned knowledge to guide its sampling, can overcome such intrinsic limitations naturally. Learning strategy, which decides the representation, selection, and usage of knowledge, may affect the performance of MSA algorithms significantly. Using the current population as learned knowledge, we design three different knowledge selection schemes, selecting from better agents, selecting from worse agents and selecting from all agents randomly, to select knowledge to guide sampling. A differential perturbation operator is designed to generate candidate solution from the selected knowledge. Comparison was carried on four widely used benchmark functions, and the results show that learning-based MSA algorithm has good performance in terms of convergence speed and solution accuracy. Furthermore, simulation results also show that even learning from worse agents significantly outperforms not learning at all.