Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Discovering analogous structure supports generalization in human reinforcement learning

Alana Jaskir¹, Michael Frank²; ¹New York University, ²Brown University

Presenter: Alana Jaskir

Humans are remarkably skilled at transferring knowledge between experiences. Reinforcement learning models, while adept at capturing human behavior in stationary environments, can struggle to adapt when the optimal policy changes substantially. Reward-predictive representations (RPRs) build abstractions by identifying states and actions with analogous long-term reward sequences, affording greater robustness to policy changes. Here, we test empirical predictions of RPRs. Participants played a novel sequential decision-making task, learning through trial and error how to unlock a safe's keypad to collect diamonds. In line with RPR predictions, participants showed evidence of learning an abstraction that they leveraged to support generalization when the key combination lock abruptly changed.

Topic Area: Memory, Learning & Knowledge Structures