Promoting Constructive Interaction and Moral Behaviors Using Adaptive Empathetic Learning

Moral system assists people with constructive interaction by maximizing the inner stimulus transfered from outer feelings. For this reason, building an intrinsic sense of morality is one potential way of regulating agents' behaviors. Incorporating ideas found in social neuroscience, we hardwired a theoretical model of empathy in rational reinforcement learning-based agents to enable affective state sharing between agents. Our learning algorithm accounts for the impact of social comparison and companion impression, which play an important role on the update of empathy and make it possible for agents to change between cooperation and competition adaptively. Empathetic learners' behavioral dynamics were tested and analyzed in multiple game settings. In iterated prisoner dilemma, empathetic agents showed increased cooperation in most cases except exhibiting self-protection awareness vigilantly when their partners were in the antagonistic state. Empathetic agents also showed a strong sense of fairness in the ultimatum game which resulted in an evenhanded allocation scheme on resources.