Evolutionary optimization in Markov random field modeling

Global optimization is a crucial and challenging problem in Markov random field modeling. In this paper we propose an evolutionary algorithm which guides the exploration of search space by building probabilistic model of promising solutions. New population is not generated using genetic operators of crossover and mutation, but sampled directly from the estimated distributions encoded in the probabilistic model. Under the selective pressure impressed by the fitness-weighted distribution estimation, population evolves generation by generation towards the global optimum. Experimental comparisons show that our algorithm outperforms genetic algorithm in both convergence speed and solution quality.