[4E1-GS-2] Machine learning: agents

[4E1-GS-2-02] Visualizing tactics of reinforcement learning agents through t-SNE dimensionality reduction in state space

〇yuki nagatomo¹, Youichiro Miyake¹ (1. Rikkyo University )

Keywords:Reinforcement Learning, Agents, Explainability

We investigated the explainability of reinforcement learning by visualizing the tactics taken by agents in re- inforcement learning. Reinforcement-learning agents are generally black boxes, and it is unclear what kind of decisions they make and what actions they take. However, by observing the transitions in the agent’s state space, we can find a pattern that leads to a series of actions. It is not easy to analyze how the patterns are formed by the innumerous state variables that exist in space-time, because of the curse of dimensionality. In order to analyze the tactics, we take the trajectories of the agent in the 2D space and analyze them by dimensionality reduction using t-SNE. As the result, we succeeded in visualizing that the agent repeatedly uses some pattern.