EXplainable Reinforcement Learning Using Introspection in a Competitive Scenario

Reinforcement learning (RL) is inspired by behavioral psychology and helps solve problems where there is no previous data; that is, the agent learns through trial and error by interacting with the environment. Explainable RL aims to solve problems related to trust and transparency that people without technical knowledge might have about these systems. This work proposes a novel explainable reinforcement learning approach based on introspection in Deep Q-network and Proximal Policy Optimization algorithms. The integration of the introspection method empowers RL agents to assess the probability of success in a game, solely based on the Q-values obtained. In this regard, the agent will be able to measure how high the chance of winning for each available action during the game using the value function approximation's output. Finally, the introspection-based agents could win several rounds during training, being more competitive than their opponents in different game moments. The computed probabilities of success, showed that although the agent was able to complete a reasonable number of games and generated strategies to win, the agent could not maintain a constant rhythm and learning process.