Comparison With Other ProposalsTopAvoiding SplittingEmpirical Results

Empirical Results

An interesting question is whether there are real examples where the advantage of exploiting contextual independence outweighs the overhead of maintaining confactors.

We can easily generate synthetic examples where VE is exponentially worse than contextual variable elimination (for example, consider a single variable with n, otherwise unconnected, parents, where the decision tree for the variable only has one instance of each parent variable, and we eliminate from the leaves). At another extreme, where all contexts are empty, we get VE with very little overhead. However, if there is a little bit of CSI, it is possible that we need to have the overhead of reasoning about variables in the contexts, but get no additional savings. The role of the empirical results is to investigate whether it is ever worthwhile trying to exploit context-specific independence, and what features of the problem lead to more efficient inference.


David Poole and Nevin Lianwen Zhang,Exploiting Contextual Independence In Probabilistic Inference, Journal of Artificial Intelligence Research, 18, 2003, 263-313.

Comparison With Other ProposalsTopAvoiding SplittingEmpirical Results