Expected scalarised returns dominance: a new solution concept for multi-objective decision making

Abstract In many real-world scenarios, the utility of a user is derived from single execution policy. this case, to apply multi-objective reinforcement learning, expected returns must be optimised. Various scenarios exist where user’s preferences over objectives (also known as function) are unknown or difficult specify. such set optimal policies learned. However, settings maximised have been largely overlooked by learning community and, consequence, solutions has yet defined. work, we propose first-order stochastic dominance criterion build solution sets maximise utility. We also define new criterion, scalarised (ESR) dominance, that extends allow learned in practice. Additionally, concept called ESR set, which dominant. Finally, present tabular distributional (MOTDRL) algorithm learn multi-armed bandit settings.