人类群体行为的研究对于探求人类行为复杂性的内在机理，发掘行为背后的规律有着重要的意义。通过行为实验和数学统计相结合的方法是研究人类行为的重要途径之一。人类群体行为中，公平和理性问题的研究有着非常重要的意义，对于探求人类社会中的合作，利他等行为有着重要的意义。最后通牒博弈是研究公平和理性问题的基本博弈模型。在最后通牒博弈中，提议者提出如何分配，回应者选择是否接受分配方案，若接受，则按照提议者的分配方案进行分配，否者双方都无法得到金钱。在本文中，我们通过一个个体同时扮演提议者和回应者的办法，消除了在最后通牒博弈中个体地位不对等的问题。我们发现，在群体双角色最后通牒博弈中，回应者的报价行为趋于理性，而提议者根据回应者的报价行采用最优反应策略调整自己的出价。在实验中，大部分个体行为采用了利他策略p>q。另外，我们发现群体平均收益跟轮次呈现正相关，而收益的方差跟轮次呈现负相关，说明群体内存在一定的协调行为。最后，我们基于强化学习建立了演化模型，对提议者的行为进行了模拟分析。

The study of human group behavior is of great significance to explore the intrinsic mechanisms of human behavior complexity and the law behind the behavior. One of the important ways to study human behavior is based on the combination of behavioral experiment and mathematical statistics. In human society, fairness and rationality are two fundamental characteristics, which also have closed relationships with human cooperation and altruism. The ultimatum game is a basic game model to study fairness and rationality. In the ultimatum game, two players, a proposer and a responder, together decide how to divide a sum of money. The proposer proposes how to distribute the money and the responder can either accept or reject it. Rejection causes both players to get nothing. In this thesis, one individual plays proposer and responder spontaneously to eliminate the influence of unequal status in the ultimatum game. We find that acceptance levels of responders approach rationality and proposers take a best-response strategy to choose their offers in our dual-role ultimatum game experiments. By classification analysis, we find most of individuals adopt altruistic strategy p>q. Furthermore, we observe positive correlations between average payoff and rounds, and negative correlations between standard deviations of payoff and rounds, which indicate there are coordinating decisions in our experiments. Finally, we establish reinforcement learning model to reproduce the evolution behaviors of proposers.