VISITOR: Visual Interactive State Sequence Exploration for Reinforcement Learning

Abstract

Understanding the behavior of deep reinforcement learning agents is a crucial requirement throughout their development. Existing work has addressed the identification of observable behavioral patterns in state sequences or analysis of isolated internal representations; however, the overall decision-making of deep-learning RL agents remains opaque. To tackle this, we present VISITOR, a visual analytics system enabling the analysis of entire state sequences, the diagnosis of singular predictions, and the comparison between agents. A sequence embedding view enables the multiscale analysis of state sequences, utilizing custom embedding techniques for a stable spatialization of the observations and internal states. We provide multiple layers: (1) a state space embedding, highlighting different groups of states inside the state-action sequences, (2) a trajectory view, emphasizing decision points, (3) a network activation mapping, visualizing the relationship between observations and network activations, (4) a transition embedding, enabling the analysis of state-to-state transitions. The embedding view is accompanied by an interactive reward view that captures the temporal development of metrics, which can be linked directly to states in the embedding. Lastly, a model list allows for the quick comparison of models across multiple metrics. Annotations can be exported to communicate results to different audiences. Our two-stage evaluation with eight experts confirms the effectiveness in identifying states of interest, comparing the quality of policies, and reasoning about the internal decision-making processes.