Finding Most Probable Explanations Using a Self-Adaptive Hybridization of Genetic Algorithms and Simulated Annealing

Bayesian belief networks (BBN’s) are a popular graphical representation for reasoning under (probabilistic) uncertainty. An important, and NP-complete, problem on BBN’s is the maximum a posteriori (MAP) assignment problem, in which the goal is find the network assignment with highest conditional probability given a set of observances, or evidence. In this paper, we present an adaptive hybrid technique combining genetic algorithms and simulated annealing, and apply it to a layered 70-node BBN. Simulated annealing is used as a type of mutation within the framework of the genetic algorithm, and the parameters of the annealing schedule are themselves adapted by the genetic algorithm. Key–Words: reasoning under uncertainty, evolutionary computation, optimization, diagnosis, belief networks.