- Keywords: Trust, Confidence, Q-learning, Reward Circuit
- Abstract: Existing works to compute trust as a numerical value mainly rely on ranking, rating or assessments of agents by other agents. However, the concept of trust is manifold, and should not be limited to reputation. Recent research in neuroscience converges with Berg's hypothesis in economics that trust is an encoded function in the human brain. Based on this new assumption, we propose an approach where a trust level is learned by an overlay of any model-free off-policy reinforcement learning algorithm. The main issues were i) to use recent findings on dopaminergic system and reward circuit to simulate trust, ii) to assess our model with reliable and unbiased real life models. In this work, we address these problems by extending Q-Learning to trust evaluation, and comparing our results to a social science case study. Our main contributions are threefold. (1) We model the trust-decision making process with a reinforcement learning algorithm. (2) We propose a dynamic reinforcement of the trust reward inspired by recent findings of neuroscience. (3) We propose a method to explore and exploit the trust space. The experiments reveal that it is possible to find a set of hyperparameters of our algorithm to reproduce recent findings on overconfidence effect in social psychology research.
- One-sentence Summary: By extending standard Q-learning to recent findings on reward circuits, we were able to develop an algorithm to mimic human trust.