Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Emergent Reciprocity Through Temporal Credit Assignment in Reinforcement Learning Agents

Le Thuy Duong Nguyen¹, Dane Malenfant², Blake Aaron Richards³; ¹Mila - Quebec Artificial Intelligence Institute, ²McGill University, McGill University, ³Google

Presenter: Le Thuy Duong Nguyen

Humans and animals have complex social and cultural systems that can span large distances and times. In North America, Plains Indigenous nations practiced reciprocity through Manitokan, effigies placed at fixed locations, where leaving surplus goods was seen as cooperative acts of resource sharing or caching. In multi-agent reinforcement learning (MARL), reciprocity has largely been defined as an emergent property through tit-for-tat policies or reputation scores that establish social norms. These perspectives fail to consider the temporal structure of rewards or the criticality of certain actions necessary for success. We present a novel MARL environment to investigate the emergence of reciprocal prosocial behaviours in reinforcement learning agents. Baseline experiments show that agents consistently converged to suboptimal policies favoring individual resource maximization, despite the potential for improved collective outcomes. These findings highlight a critical gap in existing MARL methods, suggesting the need for new algorithms capable of supporting temporal credit assignment in artificial agents.

Topic Area: Reward, Value & Social Decision Making

Extended Abstract: Full Text PDF