Evaluation of Strategies for Co-evolutionary Genetic Algorithms: Dlcga Case Study

Dafo, a multi-agent framework dedicated to distributed coevolutionary genetic algorithms (CGAs) is used to evaluate dLCGA, a new dynamic competitive coevolutionary genetic algorithm. We compare the performance of dLCGA to other known classes of CGAs for the Inventory Control Parameter optimization problem (ICP) and in particular show how it improves the results of the static version of LCGA. INTRODUCTION With the increasing complexity of real world problems considered in Computer Science, multi-agent paradigm has emerged as an efficient way to analyze and implement solutions. This paradigm offers a natural way of analyzing a problem as a distributed and decentralized generation of the solution using autonomous agents having each an aspect of the global problem to solve. This characteristic drives the increasing use of those technologies within evolutionary computation where agents are equipped with genetic algorithms. Going a step further, in this paper we are interested in using the multi-agent paradigm as a tool for implementing coevolutionary algorithms where agents classically communicate in a cooperative or competitive way, using a fixed strategy. However its appears that when dealing with real world and consequently dynamic problems, this strategy should also be dynamic in order to provide a good solution at any time: evolutionary algorithms must adapt to dynamic solution spaces. That is why our researches provide autonomy to the components of the evolutionary algorithms that consequently are capable of selfreorganization. dLCGA is a new dynamic competitive coevolutionary genetic algorithm resulting of this research. In the next section we introduce the concept of coevolutionary genetic algorithms, both cooperative and competitive, including our new dynamic variant. Section III presents the architecture of Dafo and section IV provides some comparison of our dynamic CGA to two static CGAs on an inventory management problem. Finally general conclusions are drawn and some perspectives are given. COEVOLUTIONARY ALGORITHMS As for the "classical" genetic algorithm, the concept of coevolutionary algorithms comes from biological observations (Paredis, 1996). Indeed, the nature is composed of several species that coevolve, and instead of evolving a