Annealed Importance Sampling for英文资料.pdf

Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence Annealed Importance Sampling for Structure Learning in Bayesian Networks∗ ¨ Teppo Niinimaki and Mikko Koivisto HIIT & Department of Computer Science, University of Helsinki, Finland teppo.niinimaki@cs.helsinki.fi, mikko.koivisto@cs.helsinki.fi ] Abstract 2000 , and it has been applied in various forms also to struc- ture learning in BNs. We summarize some corner stones We present a new sampling approach to Bayesian of the developments. Madigan and York [1995] presented a learning of the Bayesian network structure. Like Markov chain that moves in the space of DAGs by simple arc some earlier sampling methods, we sample linear changes. Friedman and Koller [2003] obtained a significantly orders on nodes rather than directed acyclic graphs faster-mixing chain by operating, not directly on DAGs, but (DAGs). The key difference is that we replace the in the much smaller and smoother space of node orderings. usual Markov chain Monte Carlo (MCMC) method A drawback of the sampler, order-MCMC in the sequel, is by the method of annealed importance sampling that it introduces a bias favoring DAGs that are compatible (AIS). We show that AIS is not only competitive