Preventing Premature Convergence via Cooperating Genetic Algorithms

The definition of the hardness of a problem for GA’s has been tackled, eventually leading to the notion of deception [Gol89, HG94, Dav87]. It has been known for a while that the hardness of a problem is inherently related to the representation that is used. This fact will be illustrated below by showing that an easy problem (1’s counting problem) can become nearly unsolvable after a change of representation. Then, we show how a set of cooperating GA’s, each solving the same problem but using different representations, can succeed in solving a difficult problem by exchanging individuals. Remapping the research space with a change of representation is not a new idea [BV90, MW92, KD95]. Remapping may involve an increase or a decrease of the number of local optima by dynamically changing the neighborhood of individuals. In this paper, we show how cooperation can help finding the global optimum. The idea is that when a GA is stuck on a point (because of the representation it uses), it is likely that an other GA, using an other representation provides it with new individuals from which the search can go further. Our point is not to use parallelism to speed up the computation but to use the parallelism as a new way of exploring the research space.