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

Shortest-Path Planning on Learned Latent-State Maps Better Fits Biological Navigation Data Than Tabular Successor Representations

Zacharie Bugaud¹, Yicong Zheng¹, Dileep George²; ¹Astera Institute, ²Vicarious AI

Presenter: Zacharie Bugaud

Model-based (MB) and successor representation (SR) planning are leading accounts of spatial navigation. De Cothi et al. (2022) compared tabular MB, SR, and model-free agents on a dynamic-barrier task and found that SR better predicted human choices, but all agents were given oracle state identity. We ask whether this comparison changes when agents must learn their own state representations from aliased observations. We equip both planners with Clone-Structured Cognitive Graphs (CSCGs), which infer latent states from observation-action history, and evaluate them on the same task using the original metrics. CSCG+MB achieves 93.6% goal-reaching, close to humans (94.9%), whereas CSCG+SR reaches 90.2%. CSCG+MB also better predicts human action choices (-1.01 log-lik/step vs. -1.26 for SR), reversing the action-likelihood advantage previously reported for oracle SR. SR retains a higher difficulty-pattern correlation, but this may reflect shared failure modes. Under realistic perceptual constraints, shortest-path planning on learned state maps provides a better overall account of biological navigation in this task.

Topic Area: Memory, Learning & Knowledge Structures