Causality in Multi-Agent Systems

Over the last couple of years, causality has become of bigger interest to the AI community. It has, among other things, been used to generate explanations of black-box models. Despite this interest, research into causality in strategic multi-agent systems settings has been lacking. This project intends to develop methods to study causality in multi-agent systems, with the goal of determining accountability of system outcomes. In order to do this, we first discuss what we understand by causality. We then introduce a first attempt at developing a causal model for a strategic multi-agent setting. Finally, we discuss how causal questions could be answered more efficiently using abstraction techniques.