Evolutionary learning function approximator as robot controller

In many cases, autonomous robot are required to make decisions repeatedly to select the best action plan among a set of alternatives just according to their indexes of gain and cost. A general robot controller model for such task is outlined. It is found that multi-dimensional monotone functions are sufficient for the comprehensive evaluation of plans. A new Evolutionary Decision Making (EDM) approach, based on evolutionary function approximation by Genetic Algorithms, is proposed to learn the core of such decision making strategy. An application instance on virtual exploring robot controller design is given, which validate the effectiveness of the proposed approach.