Proceedings of the Annual Meeting of the Cognitive Science Society

We show that computational reinforcement learning can modelhuman decision making in the Iowa Gambling Task (IGT). TheIGT is a card game, which tests decision making under uncer-tainty. In our experiments, we found that modulating learningrate decay in Q-learning, enables the approximation of both thebehaviour of normal subjects and those who are emotionallyimpaired by ventromedial prefrontal lesions. Outcomes ob-served in impaired subjects are modeled by high learning ratedecay, while low learning rate decay replicates healthy sub-jects under otherwise identical conditions. The ventromedialprefrontal cortex has been associated with emotion based re-ward valuation, and, the value function in reinforcement learn-ing provides an analogous assessment mechanism. Thus rein-forcement learning can provide a good model for the role ofemotional reward as a modulator of the learning rate.