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

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Representation learning and choice in naturalistically complex environments

Dale Zhou¹, Shuheng Guo¹, Michael A. Yassa¹, Aaron M Bornstein¹; ¹University of California, Irvine

Presenter: Dale Zhou

Balancing exploration of uncertain options with exploiting past rewards is challenging when similar situations can produce very different outcomes. An open question is how individuals pursue rewards when the reward-maximizing strategy requires implausibly detailed memory for the landscape of potential choice options. Here, we examined how humans navigate such rugged reward landscapes under limits on representational resources. Using an information-theoretic framework, we quantified policy complexity (mutual information between states and actions) as a measure of representational cost. Participants (n=49) performed a naturalistic foraging task in a modified Super Mario environment, learning rewards across multidimensional, nonlinearly interdependent features. Behavior ranged from simple heuristics to complex state-dependent policies. Exploration induced dimensionality reduction in state representations, while greater policy complexity over selected features predicted higher reward, better memory, and improved generalization to novel stimuli. Together, these results suggest that humans dynamically balance reward maximization with representational efficiency in complex environments.

Topic Area: Decision-Making, Cognitive Control & Event Cognition