Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Uncovering brain-wide planning strategies with deep RL: Lessons from the Tower of Hanoi

Michele Garibbo¹, Austin Tudor David Andrews², Jascha Achterberg², Rui Ponte Costa²; ¹University of Bristol, ²University of Oxford

Presenter: Austin Tudor David Andrews

Human planning involves generating and executing ac- tion sequences under environmental constraints (Mattar & Lengyel, 2022). Experimental studies have identified that areas such as the prefrontal cortex (PFC) , hippocam- pus and cerebellum play important roles during planning (Grafman et al., 1992; Goel & Grafman, 1995). We pro- pose that the architectures of deep reinforcement learn- ing agents capable of solving human-level planning tasks can offer a normative framework for understanding the involvement of different brain regions in planning. To demonstrate this, we use MuZero (Schrittwieser et al., 2020) and a widely used task to study goal-directed plan- ning and behavior, Tower-of-Hanoi (ToH). We evaluate the performance of MuZero on the ToH under targeted net- work ablations to simulate brain region-specific lesion studies. Ablating the value network reproduces the be- havior observed in patients with PFC damage, while ab- lating the policy network mimics cerebellar damage. Our preliminary results suggest that deep RL architectures may provide a brain-wide account of human planning.

Topic Area: Reward, Value & Social Decision Making

Extended Abstract: Full Text PDF