On fairness and learning agents in a bargaining model with uncertainty
详细信息 查看全文
- 作者：Nawa ; Norberto Eiji ; Shimohara ; Katsunori ; Katai ; Osamu
- 关键词：Computational agents ; Alternating-offers bargaining model ; Fairness ; Agent-based modeling ; Q-learning
- 刊名：Cognitive Systems Research
- 出版年：2002
- 出版时间：December, 2002
- 年：2002
- 卷：3
- 期：4
- 页码：555-578
- 全文大小：429 K
文摘
The emergence of fair divisions in a repeated bargaining game is investigated in a computational model. Individuals of human societies often appeal to some norm of fairness in situations where an agreement over the division of a surplus is required. The employed framework consists of players alternating offers that describe possible ways to share a certain commodity. The players are allowed a limited number of offers to reach an agreement; if they fail to agree, the player who made the first offer, the lucky player, wins the whole lot at stake. Uncertainty is introduced in the process by randomly choosing the lucky player at the beginning of each iteration. In the experiments, the players acquired strategies by employing a variant of Q-learning, a reinforcement learning algorithm. Experiments were performed with different configurations of utility functions on the players’ preferences in taking actions in risky situations. Analysis of the results shows that the game theoretical model of a single shot of the bargaining game used in the experiments closely matches the outcomes obtained in the simulated framework, despite the differences in the quality of the players, who are assumed to be fully rational in the theoretical model. Learning agents that are timid toward risky situations manage to acquire strategies that lead to fair outcomes when playing against each other, but find themselves in a disadvantageous position when confronting bolder types.