Controlling Exploration , Diversity and Escaping Local Optima in GP : Adapting Weights of Training Sets to Model Resource Consumption

A common problem in evolutionary computation, and in particular in GA and GP, is the loss of diversity due both to ‘lock-in’ and early convergence of a population to ‘deceptive’ high scoring partial solutions. As the run proceeds, such individuals constitute a larger and larger segment of the population, eliminate diversity and stop the progress of the run. This paper will present a method for the automatic detection of such lock-in and impasse, as well as explore a fitness function which uses this information to reward diversity and innovation and increasingly penalize the locked-in individuals. This method can be applied to essentially any problem, and can be successful in preventing a run from locking-in on an easily obtainable local optimum and in preserving diversity. We present an analysis, discussion, and preliminary results on its application to the boolean 11-multiplexer problem.