| Popis: |
Reinforcement Learning (RL) has emerged as a powerful tool to solve sequential decision-making problems, where a learning agent interacts with an unknown environment in order to maximize its rewards. Although most RL real-world applications involve multiple agents, the Multi-Agent Reinforcement Learning (MARL) framework is still poorly understood from a theoretical point of view. In this manuscript, we take a step toward solving this problem, providing theoretically sound algorithms for three RL sub-problems with multiple agents: Inverse Reinforcement Learning (IRL), online learning in MARL, and policy optimization in MARL. We start by considering the IRL problem, providing novel algorithms in two different settings: the first considers how to recover and cluster the intentions of a set of agents given demonstrations of near-optimal behavior; the second aims at inferring the reward function optimized by an agent while observing its actual learning process. Then, we consider online learning in MARL. We showed how the presence of other agents can increase the hardness of the problem while proposing statistically efficient algorithms in two settings: Non-cooperative Configurable Markov Decision Processes and Turn-based Markov Games. As the third sub-problem, we study MARL from an optimization viewpoint, showing the difficulties that arise from multiple function optimization problems and providing a novel algorithm for this scenario. |
