Conditioning Methods for Exact and Approximate Inference in Causal Networks
This work addresses inference challenges in causal networks, offering improved efficiency and flexibility, but it appears incremental as it refines existing methods like cutset conditioning.
The paper tackles the problem of exact and approximate inference in causal networks by introducing two algorithms: dynamic conditioning, which reduces complexity from exponential to linear in some cases, and B-conditioning, which enables trade-offs between approximation quality and computation time.
We present two algorithms for exact and approximate inference in causal networks. The first algorithm, dynamic conditioning, is a refinement of cutset conditioning that has linear complexity on some networks for which cutset conditioning is exponential. The second algorithm, B-conditioning, is an algorithm for approximate inference that allows one to trade-off the quality of approximations with the computation time. We also present some experimental results illustrating the properties of the proposed algorithms.