Altruism and Selfishness in Believable Game Agents: Deep Reinforcement Learning in Modified Dictator Games

Abstract

This article focuses on using deep reinforcement learning, specifically proximity policy optimization, to train agents in a social dilemma game, modified dictator game, in order to investigate the effect of selfishness and altruism on the believability of the game agents. We present the design and implementation of the training environment, including the reward functions which are based on the findings of established empirical research, with three agent profiles mapped to the three standard constant elasticity of substitution (CES) utility functions, i.e., selfish, perfect substitutes, and Leontief, which measure different levels of selfishness/altruism. The trained models are validated and then used in a sample game, which is used to evaluate the believability of the three agent profiles using the agent believability metrics. The results indicate that players find altruistic behaviour more believable and consider selfishness less so. Analysis of the results indicates that human-like behavior resulting from the application of artificial intelligence evolves from perceived human behavior rather than the observed. The analysis also indicates that selfishness/altruism may be considered as an extra dimension to be included in the believability metrics.