DTAS: An Adaptive Critical Scenarios Generation Method for Decision Boundary Assessment

Intelligent algorithms have been widely applied in autonomous decision-making systems (ADMSs). Despite the success of these technologies in practical applications, our understanding of the underlying mechanisms of system decisions remains limited. A key factor in studying the decision of ADMSs is the decision boundary (DB), which helps us understand the decision behavior and reflects the safety-critical aspects and decision robustness of the system. However, this depends on the ability to generate critical scenarios near the system's DB. To address this issue, this study proposes a critical boundary scenarios (CBSs) generation method called decision tree-assisted adaptive sampling (DTAS), these CBSs effectively cover the DB of the ADMSs, thereby fully characterizing the DB. In this study, we focus solely on the input and output of the system, treating the ADMS as a completely black-box, making DTAS suitable to any black-box ADMSs with discrete outputs. Additionally, we propose a local DB description method based on decision rule optimization. These decision rules improve the interpretability of complex DB by describing parameter regions. The proposed methods are conducted extensive experiments on standard benchmarks and actual autonomous systems. The experiment performance exhibits the effectiveness of our approaches.