Adaptation of a multiagent evolutionary algorithm to NK landscapes

Abstract

Multiagent evolutionary algorithm (MEA) is a relatively new optimization technique, where a life cycle of a population of agents, which perform local search, is simulated. The algorithm was originally intended as a method for solving the graph coloring problem and incorporates ideas such as lifespans of agents and a positive or negative reinforcement for the ability of the agent to improve fitness or its stagnation. In this paper, we propose to use MEA for optimization on NK fitness landscapes. These landscapes are popular for the tunability of their ruggedness and are a particularly interesting use case for MEA. This algorithm is especially well suited for functions, where local search tends to fail because of their multimodality. However, using many short-term local search subroutines in a well-tuned version of MEA can significantly improve the results of the same local search algorithm. Experimental results are presented for MEA with the simple (1+1) Evolutionary Algorithm ((1+1) EA) used as a local search subroutine. These results show that in large and more rugged NK landscapes, MEA outperforms the multi-start (1+1) EA with number of parallel starts equal to the initial population size of MEA. This is the first time we obtained results, which clearly indicate that solely the emergent multiagent nature of MEA, driven by the lifespans and the reinforcement mechanism, is able to improve the results of multi-start local search.