In this paper, we explore semantic clustering properties of deep reinforcement learning (DRL) to improve its interpretability and deepen our understanding of its internal semantic organization. In this context, semantic clustering refers to the ability of neural networks to cluster inputs based on their semantic similarity in the feature space. We propose a DRL architecture that incorporates a novel semantic clustering module that combines feature dimensionality reduction with online clustering. This module integrates seamlessly into the DRL training pipeline, addressing the instability of t-SNE and eliminating the need for extensive manual annotation inherent to prior semantic analysis methods. We experimentally validate the effectiveness of the proposed module and demonstrate its ability to reveal semantic clustering properties within DRL. Furthermore, we introduce new analytical methods based on these properties to provide insights into the hierarchical structure of policies and semantic organization within the feature space. Our code is available at https://github.com/ualiangzhang/semantic_rl.

Compound graphs are networks in which vertices can be grouped into larger subsets, with these subsets capable of further grouping, resulting in a nesting that can be many levels deep. In several applications, including biological workflows, chemical equations, and computational data flow analysis, these graphs often exhibit a tree-like nesting structure, where sibling clusters are disjoint. Common compound graph layouts prioritize the lowest level of the grouping, down to the individual ungrouped vertices, which can make the higher level grouped structures more difficult to discern, especially in deeply nested networks. Leveraging the additional structure of the tree-like nesting, we contribute an overview+detail layout for this class of compound graphs that preserves the saliency of the higher level network structure when groups are expanded to show internal nested structure. Our layout draws inner structures adjacent to their parents, using a modified tree layout to place substructures. We describe our algorithm and then present case studies demonstrating the layout's utility to a domain expert working on data flow analysis. Finally, we discuss network parameters and analysis situations in which our layout is well suited.