A recurrent network approach to MAP explanation

Bayesian belief networks (BBNs) are an increasingly popular knowledge representation for reasoning under (probabilistic) uncertainty. An important problem in BBNs is finding the best, i.e. the most probable, explanation for a given set of observations, called the evidence. In this paper, we present a recurrent neural network approach to the maximum a-posteriori (MAP) problem. We measure the performance of our approach on more than 300 pairs of belief networks and evidence sets: a combination of 23 different networks and between 10 and 21 evidence sets on each network. We find that, on average, our neural network is able to return solutions within 95% of the probability of the optimal solution