Evolutionary Multi-Agent System in Hard Benchmark Continuous Optimisation

Abstract

It turns out that hybridizing agent-based paradigm with evolutionary computation brings a new quality to the field of meta-heuristics, enhancing individuals with possibilities of perception, interaction with other individuals (agents), adaptation of parameters, etc. In the paper such technique--an evolutionary multi-agent system (EMAS)--is compared with a classical evolutionary algorithm (Michalewicz model) implemented with allopatric speciation (island model). Both algorithms are applied to the problem of continuous optimisation in selected benchmark problems. The results are very promising, as agent-based computing turns out to be more effective than classical one, especially in difficult benchmark problems, such as high-dimensional Rastrigin function.