Evolution of Subjective Utilities by GA with BLX-α

Abstract

Utility-based Q-learning, which uses subjective utilities as rewards of Q-learning, has been proposed and the utilities that derive mutual cooperation in a Prisoner's Dilemma game have been successfully evolved by real-coded genetic algorithm (RCGA). However, in that work, the genes were simply exchanged in the evolution process like a bit-string GA and the search space was not so wide as a result. This work investigates the evolution of the subjective utilities by RCGA with blend crossover (BLX-α) that has a powerful search ability by generating various chromosomes.