Decision-making in team-reward multi-objective multi-agent domains

The majority of multi-agent system implementations aim to optimise
agents’ policies with respect to a single objective, despite the fact
that many real-world problem domains are inherently multi-objective
in nature (e.g., mitigation strategies in epidemiological modelling,
electricity grid management). Multi-objective multi-agent systems are
a more general framework that explicitly consider the possible tradeoffs
between conflicting objective functions. Up until recently,
research in this highly complex domain was sparse and fragmented.
For this project we pursue to push forward the developments in this
field, further bridging the gap between learning systems and realworld
problem settings. To this end, our goal is to develop
reinforcement learning approaches for sequential team-reward multiobjective
multi-agent systems. More precisely, we consider the case
in which all the agents receive the same vectorial reward signal, but
where the final utility of the user is not known during learning.
Additionally, we also look at the case in which the agents represent
distinct users, each having a different preference over the objectives.
The outcome of our proposed methods provides decision support in
the form of a set of trade-off policies for the user, or equilibrium
policies that strike a compromise between the users’ different
preferences over the considered objectives.