Modelling Emotion Based Reward Valuation with Computational Reinforcement Learning
- Can Koluman, Department of Computer Science, City, University of London, London, United Kingdom
- Christopher Child, Department of Computer Science, City, University of London, London, United Kingdom
- Tillman Weyde, City, University of London, City, University of London, London, United Kingdom
AbstractWe show that computational reinforcement learning can model human decision making in the Iowa Gambling Task (IGT). The IGT is a card game, which tests decision making under uncertainty. In our experiments, we found that modulating learning rate decay in Q-learning, enables the approximation of both the behaviour of normal subjects and those who are emotionally impaired by ventromedial prefrontal lesions. Outcomes observed in impaired subjects are modeled by high learning rate decay, while low learning rate decay replicates healthy subjects under otherwise identical conditions. The ventromedial prefrontal cortex has been associated with emotion based reward valuation, and, the value function in reinforcement learning provides an analogous assessment mechanism. Thus reinforcement learning can provide a good model for the role of emotional reward as a modulator of the learning rate.