ABSTRACT

Agent-based modelling is a powerful tool when simulating human systems, yet when human behaviour cannot be described by simple rules or maximizing one’s own profit, we quickly reach the limits of this methodology. Machine learning has the potential to bridge this gap by providing a link between what people observe and how they act in order to reach their goal. In this paper we use a framework for agent-based modelling that utilizes human values like fairness, conformity and altruism. Using this framework we simulate a public goods game and compare to experimental results. We can report good agreement between simulation and experiment and furthermore find that the presented framework outperforms strict reinforcement learning. Both the framework and the utility function are generic enough that they can be used for arbitrary systems, which makes this method a promising candidate for a foundation of a universal agent-based model.

摘要

在模拟人类系统时，基于代理的建模是一种强大的工具，但是当人类行为不能用简单的规则或最大化自己的利润来描述时，我们很快就会达到这种方法的极限。机器学习有可能通过在人们观察到的事物和他们为实现目标而采取的行动之间建立联系来弥合这一差距。在本文中，我们使用了一个基于代理的建模框架，该框架利用了公平、顺从和利他主义等人类价值观。使用这个框架，我们模拟了一个公共物品游戏并与实验结果进行比较。我们可以报告模拟和实验之间的良好一致性，并且进一步发现所提出的框架优于严格的强化学习。框架和效用函数都足够通用，可以用于任意系统，