Model-based reinforcement learning: a computational model and an fMRI study

In this paper, we discuss an optimal decision-making problem in an unknown environment on the bases of both machine learning and brain learning. We present a model-based reinforcement learning (RL) in which the environment is directly estimated. Our RL performs action selection according to the detection of environmental changes and the current value function. In a partially-observable situation, in which the environment includes unobservable state variables, our RL incorporates estimation of unobservable variables. We propose a possible functional model of our RL, focusing on the prefrontal cortex and the anterior cingulate cortex. To test the model, we conducted a human imaging study during a sequential learning task, and found significant activations in the dorsolateral prefrontal cortex and the anterior cingulate cortex during RL. From a comparison of the mean activations in the earlier and later learning phases, we suggest that the dorsolateral prefrontal cortex maintains and manipulates the environmental model, while the anterior cingulate cortex is related to the uncertainty of action selection. These experimental results are consistent with our model.