Abductive inference in Bayesian networks using distributed overlapping swarm intelligence

In this paper we propose several approximation algorithms for the problems of full and partial abductive inference in Bayesian belief networks. Full abductive inference is the problem of finding the \(k\) most probable state assignments to all non-evidence variables in the network while partial abductive inference is the problem of finding the \(k\) most probable state assignments for a subset of the non-evidence variables in the network, called the explanation set. We developed several multi-swarm algorithms based on the overlapping swarm intelligence framework to find approximate solutions to these problems. For full abductive inference a swarm is associated with each node in the network. For partial abductive inference, a swarm is associated with each node in the explanation set and each node in the Markov blankets of the explanation set variables. Each swarm learns the value assignments for the variables in the Markov blanket associated with that swarm’s node. Swarms learning state assignments for the same variable compete for inclusion in the final solution.