Evolutionary Learning Strategy using Bug-Based Search

We introduce a new approach to GA (Genetic Algorithms) based problem solving. Earlier GAs did not contain local search (i.e. hill climbing) mechanisms, which led to optimization difficulties, especially in higher dimensions. To overcome such difficulties, we introduce a "bug-based" search strategy, and implement a system called BUGS2. The ideas behind this new approach are derived from biologically realistic bug behaviors. These ideas were confirmed empirically by applying them to some optimization and computer vision problems. 1 Introduction This paper introduces a new GA (Genetic Algorithms) based approach to evolutionary learning. The purpose of our research is to apply artificial life systems to practical problem solving or to establish problem solving from nature. Traditional GAs optimize functions by adaptive combination (crossover) and mutation of coded solutions to problems (i.e. points in the problem's search space) [Goldberg89]. These genetic search mechanisms often suffer from optimization difficulties cased by premature convergence [Schaffer91] or hamming cliff [Caruana88]. Although some recent works intend to realize local search operators for GAs [Ackley87][Muhlenbein89], it is very difficult to switch gradually from global search to local search by adaptive methods. Moreover, as shown in [Rechenberg86], for search in higher dimensions typical in computer vision applications (e.g. see the 11 dimensional search space of Fig.6) there is little hope in using random search or simple recombination methods. However , even for a higher dimensional space, in general the essential search dimensions are relatively few. Thus the best way is to climb up the gradient hill in these essential dimensions. To solve these difficulties, we present an evolutionary learning system, called BUGS2. In our earlier paper, we described the implementation of BUGS, a bug-based search system using GAs [Iba92a]. The basic idea of BUGS is that the GA chromosomes represent directional control codes rather than positional vectors and that selection criteria is based on cumulative fitness. These are derived from evolutionary learning models of predatory behaviors [Iba92c]. BUGS2 extends BUGS by incorporating more biologically realistic bug behaviors, such as sexual/asexual reproductions, variable size of population , resource sharing and resource competition. In the extended strategy an analogy is made between the value (at a given point) of a function to be maximized, and the density of bacteria at that point. As a result, the usual GA adaptive search method is integrated more naturally with hill climbing search. These ideas are empirically confirmed by applying them to optimization and computer …