Norm Inference in Social Dilemmas: An Inverse Reinforcement Learning Approach

Cooperation is an important tool for humans, crucial to reach optimal and ethical behaviour in many contexts. Multi-agent Reinforcement Learning techniques are an excellent instrument for studying the emerging cooperative behaviour of AI agents in different environments that can be simulated through games, which can be considered simplifications of the real world. Some of the most studied cases are Social Dilemmas, such as the Common Pool Resource Problem, where the cooperation or defection of the agents is crucial to the outcomes of the collective and the individuals. The latest research has led to a good performance of the cooperation between AI agents. However, another critical characteristic agents need is Norm Inference, which is the ability to identify and understand the social norms that govern behaviour in a society. It is a serious aspect that must be considered when designing them since artificial learning agents will likely be embodied in our future society and will need to interact with both humans and non-human agents. In this dissertation, an Inverse Reinforcement Learning (IRL) approach is used on the problem of Norm Inference in a Common Pool Resource problem, where the norm of private areas has been established. It is shown how it is possible to recover the expert policy that follows the norm through IRL and how the recovered reward function can be informative about the norm.