Optimal Reinforcement Learning with Asymmetric Updating in Volatile Environments: a Simulation Study

The ability to predict the future is essential for decision-making and interaction with the environment to avoid punishment and gain reward. Reinforcement learning algorithms provide a normative way for interactive learning, especially in volatile environments. The optimal strategy for the classic reinforcement learning model is to increase the learning rate as volatility increases. Inspired by optimistic bias in humans, an alternative reinforcement learning model has been developed by adding a punishment learning rate to the classic reinforcement learning model. In this study, we aim to 1) compare the performance of these two models in interaction with different environments, and 2) find optimal parameters for the models. Our simulations indicate that having two different learning rates for rewards and punishments increases performance in a volatile environment. Investigation of the optimal parameters shows that in almost all environments, having a higher reward learning rate compared to the punishment learning rate is beneficial for achieving higher performance which in this case is the accumulation of more rewards. Our results suggest that to achieve high performance, we need a shorter memory window for recent rewards and a longer memory window for punishments. This is consistent with optimistic bias in human behavior.