Abstract
Various methods of improving the validation performance of behavior selection systems for autonomous robots, based on the utility function (UF) method is investigated by means of simulations involving a robot carrying out a simple exploration task. In the UF method, behavior selection is based on utility functions which, in turn, are optimized using an evolutionary algorithm (EA). The computer simulations used for generating the utility functions can be set up in many different ways and, in this paper, several different setups are compared with respect to their ability of generating reliable behavior selection systems. The results show that better validation performance is obtained in cases where several simulations are carried out in the evaluation of an individual, provided that the simulations are sufficiently long so that the robot will have to make many behavior selection decisions. Furthermore, the results indicate that the choice of method for combining the results obtained in different simulations to form a scalar fitness value has only a rather limited impact on the performance.
