Optimized Test Scenario Identification Method based on Refined Warshall's Dynamic Programming for Validating Reinforcement Learning Model

초록

In recent decades, Reinforcement learning (RL) has achieved remarkable performance in sequential decision-making tasks; however, validating RL-based software remains a significant challenge due to the exponential growth in the number of state–action combinations. We propose an optimized test-scenario identification method for a formal validation mechanism that integrates a refined Warshall's dynamic programming algorithm and path optimization to ensure efficient test coverage for RL systems. The method constructs a directed abstract graph from the RL model, applies transitive closure analysis to check reachability between states, and identifies missing states/transitions before testing. Our goal is to use test-scenario optimization to generate a minimal yet sufficient set of test cases, thereby achieving maximum coverage with minimal redundancy. This approach reduces verification complexity while maintaining mathematical rigor, making it well-suited for safety-critical applications such as autonomous driving. The proposed mechanism provides a scalable, interpretable validation process, offering a foundation for the reliable deployment of RL–based software in real-world systems. © 2026 KSII.

키워드