On Modeling Human Learning in Sequential Games with Delayed Reinforcements

We model human learning in a repeated and sequential game context that provides delayed reinforcements. Our context is signifcantly more complex than previous work in behavioral game theory, which has predominantly focused on repeated single-shot games where the actions of other agent are perfectly observable and provides for an immediate reinforcement. In this complex context, we explore several established reinforcement learning models including temporal difference learning, SARSA and Q-learning. We generalize the default models by introducing behavioral factors that are refective of the cognitive biases observed in human play. We evaluate the model on data gathered from new experiments involving human participants making judgments under uncertainty in a repeated strategic and sequential game. We analyze the descriptive models against their default counterparts and show that modeling human aspects in reinforcement learning signifcantly improves predictive capabilities. This is useful in open and mixed networks of agent and human decision makers.