Reinforcement Learning in Multi-Objective Multi-Agent Systems: Doctoral Consortium

For effective decision-making in the real world, artificial agents need to take both the multi-agent as well as multi-objective nature of their environments into account. These environments are formalised as multi-objective games and introduce numerous challenges compared to their single-objective counterpart. For my main contributions so far, I have established a theoretical guarantee that a bidirectional link always exists that maps a finite multi-objective game to an equivalent single-objective game with an infinite number of actions. Additionally, I presented an extensive study of Nash equilibria in multi-objective games, culminating in existence guarantees under certain assumptions. From a reinforcement learning perspective, I explored how communication and commitment can help agents to learn adequate policies in these challenging environments. In this paper, I summarise my ongoing research and discuss several promising directions for future work.