Constraint-based Co-evolutionary Genetic Programming for Bargaining Problems

This thesis applies evolutionary algorithms to tackle bargaining games. Evolutionary algorithms can discover efficient and stationary strategies for various bargaining games. Game-theoretic method requires a substantial amount of mathematical reasoning. Thus this method restricts to simple problems. Moreover, game-theoretic solutions rest on the crucial assumption that every player is perfectly rational. These characteristics cast doubts on the applications of game-theoretic method to complex bargaining problems. To overcome such limitations of game-theoretic method, we adopt an alternative method, evolutionary algorithms. We assume that players are boundedly rational. We present a theoretic framework on the basis of co-evolutionary algorithms. We develop Constraint-based Co-evolutionary Genetic Programming system (CCGP) to simulate seven types of two-player bargaining scenarios. On the ground of experimental observations, the co-evolutionary algorithm successfully discovers satisfied and profitable solutions. The computational cost and human efforts of using the co-evolutionary algorithm for bargaining problems are affordable. The CCGP system is reusable. In particular, this thesis • simulates boundedly rational players’ adaptive learning in two-player bargaining games; • investigates fitness evaluations in co-evolutionary systems; • presents a constraint handling technique integrated into evolutionary algorithms. This technique is able to handle situations where both hard and soft constraints exist;