Is the Genetic Algorithm a Cooperative Learner?

This paper begins to explore an analogy between the usual competitive learning metaphor presented in the GA literature and the cooperative learning metaphor discussed by Clearwater, et al. Examining the GA in the cooperative learning framework provides additional insight into the power of the algorithm. The illustrative empirical study reported in this paper demonstrates the importance of GA parameter settings in viewing the GA as a cooperative learner. The study also points to some potential avenues of research in improving the performance of the GA. 1.0 Introduction The classical analyses of the standard genetic algorithm (GA) (Holland, 1975; De Jong, 1975) focus on two fundamental concepts: intrinsic parallelism and the formation of building blocks. In this conceptual framework, a population represents a set of competing structures that explore various parts of the search space in parallel. During each generation, the algorithm determines what parts of the search space to explore: first, by applying the selection and reproduction operators to generate clones, then by applying the crossover operators to generate new structures, and finally, by mutating these new structures by a small amount to ensure against premature convergence to a sub-optimal solution. The crossover operator builds new structures by using substructures of the clones as building blocks. What if, instead of emphasizing the notion of competition, we examine the GA in a complementary way? Suppose that we regard the GA population as a set of cooperative agents. The members are cooperative in that they share information; i.e., they communicate partial solutions to one another. (Notice that we are not considering the case where the environment evaluates the population as a whole based on its performance in achieving a common goal.) Clearwater et al (Huberman, 1990; Clearwater, 1991) describe a model of cooperative learning among several agents that share information through a blackboard. In this model, an agent may either search inIs the Genetic Algorithm a Cooperative Learner?