Task complexity interacts with state-space uncertainty in the arbitration process between model-based and model-free reinforcement-learning at both behavioral and neural levels

SUMMARY

A major open question concerns how the brain governs the allocation of control between two distinct strategies for learning from reinforcement: model-based and model-free reinforcement learning. While there is evidence to suggest that the reliability of the predictions of the two systems is a key variable responsible for the arbitration process, another key variable has remained relatively unexplored: the role of task complexity. By using a combination of novel task design, computational modeling, and model-based fMRI analysis, we examined the role of task complexity alongside state-space uncertainty in the arbitration process between model-based and model-free RL. We found evidence to suggest that task complexity plays a role in influencing the arbitration process alongside state-space uncertainty. Participants tended to increase model-based RL control in response to increasing task complexity. However, they resorted to model-free RL when both uncertainty and task complexity were high, suggesting that these two variables interact during the arbitration process. Computational fMRI revealed that task complexity interacts with neural representations of the reliability of the two systems in the inferior prefrontal cortex bilaterally. These findings provide insight into how the inferior prefrontal cortex negotiates the trade-off between model-based and model-free RL in the presence of uncertainty and complexity, and more generally, illustrates how the brain resolves uncertainty and complexity in dynamically changing environments.