面向查询高效验证规划的主动认知控制

In AI planning, Verified Planning aims to generate plans that not only achieve goals but also satisfy specific formal properties. However, the verification process is often computationally expensive, especially when multiple queries to an external verifier are required. This paper introduces a novel paradigm called Active Epistemic Control (AEC) to significantly enhance the query efficiency of verified planning. AEC minimizes interactions with external verifiers by actively selecting the most informative queries. The core idea is to leverage uncertainty quantification and information gain concepts to guide the planner. During plan generation, AEC prioritizes verifying properties that have the greatest impact on planning decisions and exhibit the highest uncertainty within the current epistemic state. The AEC framework is designed to be integrated with various planning algorithms and verifiers. Experimental results demonstrate that AEC can generate high-quality verified plans with substantially fewer verification queries across a range of complex planning domains, thereby significantly reducing the overall computational overhead of verified planning. This approach offers a new avenue for developing more practical and efficient reliable AI systems, addressing a critical bottleneck in the deployment of trustworthy autonomous agents.