An ACO Algorithm for the Most Probable Explanation Problem

We describe an Ant Colony Optimization (ACO) algorithm, ANT-MPE, for the most probable explanation problem in Bayesian network inference. After tuning its parameters settings, we compare ANTMPE with four other sampling and local search-based approximate algorithms: Gibbs Sampling, Forward Sampling, Multistart Hillclimbing, and Tabu Search. Experimental results on both artificial and real networks show that in general ANT-MPE outperforms all other algorithms, but on networks with unskewed distributions local search algorithms are slightly better. The result reveals the nature of ACO as a combination of both sampling and local search. It helps us to understand ACO better, and, more important, it also suggests a possible way to improve ACO.