Harvard · Kempner
Learning and social behavior in multi-agent worlds
Multi-agent environments and behavioral analysis for recurrent reinforcement-learning agents.
Supervised by Ryan Badman in Kanaka Rajan's group. Collaborative work with Riley Simmons-Edler and the team.
The question
Naturalistic foraging brings several problems together: navigating a large world, remembering resources, avoiding threats, and responding to other agents. These environments let us study behavior and internal representations within the same task.
My contribution to the collaboration
I work with Ryan Badman and Riley Simmons-Edler in Kanaka Rajan's group. My contribution focuses on the multi-agent environment: team and role mechanics, reward logic, experiment configurations, and behavioral logging. These components support the team's broader study of spatial memory and social behavior.
Building on ForageWorld
The project builds on the single-agent ForageWorld research by Simmons-Edler, Badman and colleagues, published at NeurIPS 2025. That work links behavior with recurrent neural representations in a partially observed foraging environment. Our collaboration extends this setting toward multi-agent interaction, including cooperation and competition.
The research direction
We use recurrent deep reinforcement-learning agents to investigate cooperative and competitive behavior. The broader aim is to connect behavioral analysis with the information represented in recurrent hidden states, drawing on methods from neuroscience.
Research this builds on · NeurIPS 2025
The single-agent study that this collaboration builds on. My contribution is to the ongoing multi-agent extension.