How can we audit an AI agent before it is safe and practical to deploy? A central task is to estimate its performance under the intended deployment policy (the target policy). However, historical observations may provide insufficient coverage of relevant states and actions, limiting evaluation reliability. We propose CAD (Flow-Constrained Audit Design), a framework that combines historical logs with a limited budget of targeted audit trials (probes). The key idea is to design probes that minimize the mismatch between the combined state-action visitation distribution (occupancy) of the logs and probes and that of the target policy. We formulate this design as a constrained convex optimization problem, directing probes toward regions important to the target policy but underrepresented in historical logs. We develop an adaptive algorithm that iteratively estimates environment dynamics and occupancies to update the probe policy. We link occupancy mismatch to bounds on mean squared error and asymptotic variance, and establish error bounds for CAD's adaptive estimator in finite and continuous state spaces. Experiments in Gridworld, LunarLander, and autonomous driving using real-world logs from the Waymo Open Motion Dataset show improved estimation accuracy over alternative data collection and probe designs.

翻译：暂无翻译