A Utility-Based Perspective on Multi-Objective Multi-Agent Decision Making

Numerous real-world problems involve multiple interacting entities and are inherently multi-objective in nature. Multi-objective multi-agent systems are a suitable paradigm to model such settings Despite the rising interest in this field, it has become difficult to compare or categorise approaches and identify the state-of-the-art solutions. Therefore, our first contribution is to develop a new taxonomy on the basis of the reward structures and utility functions, to offer a more structured view of the field. We note that utility functions are usually modelled as weights that define preferences over objectives, despite the fact that in many problems this assumption is not valid. We analyse the effect of non-linear utility functions on the set of equilibria in general multi-objective normal form games, under different optimisation criteria and look at how opponent modelling can aid the learning process in this setting. For future work, we are interested in how sequential settings can be approached under these considerations, to get a step closer to creating hybrid, artificial-and-human, multi-agent collectives that can deal with the different preferences w.r.t. the objectives of the different agents in the collective.