From Optimal Policies to Individual Differences: Rethinking Reinforcement Learning for Biology
Date:
Complex Anarchism Network (CAN) Seminar Series, OnlineTalk Page
List of Talks (Biweekly)
2025 Symposium (Biannual)
Related Project Page
Presenting a position/opinion on an underreported/understudied gap between reinforcement learning and biology + discussing how to extend this approach toward modeling alternative (anarchic) social system dynamics.
Abstract:
While useful for engineering optimization, classic reinforcement learning (RL) is not quite up to the task to modeling biological learning, as the approach tends toward a single “optimum” rather than the widespread variation (individual differences) we see in nature. They propose a number of ways to expand RL, two of which are particularly relevant for us: multi-agent environments and social structure. Both of which are critical in encouraging divergence (of personalities, groups, networks, classes). The talk will start with a brief overview of RL, as a rich modeling framework that builds on agent-based modeling, and then the position of the paper - in how to push RL toward affirming behavioral diversity. Expanding from here, we will host a discussion in how we, as anarchists, may use simulations (RL or not) to study alternative social system dynamics (to inform our organizational decisions + as outreach to non-anarchists).
