Causal Learning and Reasoning in Multi-Agent Reinforcement Learning

Reinforcement Learning (RL) has achieved remarkable success in sequentialdecision-makingtasks, particularlyunderrestrictiveconditions such as full observability, on-policy interaction, and online learning. However, outside these conditions, hence especially in Multi-Agent (MA) domains, policies become brittle to distribution shifts and confounding, leading to poor generalization and limited transferability. Motivated by these limitations, my PhD project aims to identify 𝑤ℎ𝑒𝑛 and 𝑤ℎ𝑦 causal learning and reasoning are necessary or desirable in (MA)RL, and ℎ𝑜𝑤 they can be systematically integrated within (MA)RL methods to improve robustness, efficiency, and interpretability of the decision-making process.

Authors

• Giovanni Briglia Department of Sciences and Methods for Engineering, University of Modena and Reggio Emilia

Keywords

• CausalReasoning
• CausalInference
• ReinforcementLearning
• Multi- Agent Reinforcement Learning
• Causal Reinforcement Learning