This paper proposes a novel cognitive planning algorithm based on Dynamic Perception Logic (DEL). Its core principle is to limit the planning agent's inference depth to an upper bound b, ensuring that it can only infer about higher-order knowledge with dimensions up to b. By iteratively increasing b, we compute the plan requiring the lowest inference depth. We utilize a novel type of "canonical" b-similarity contraction to ensure a unique minimal model by construction. This ensures smaller states and efficient tracking of visited states compared to standard similarity contraction. We prove the correctness and completeness of the planning algorithm under an appropriate inference depth bound and show that the time complexity for b is (b+1)-EXPTIME. We implement the algorithm on DAEDALUS, a novel cognitive planning tool, and demonstrate significant performance improvements over the existing EFP 2.0 planning tool on several benchmarks.
