Probabilistic Graphical Model based Approach to Genetic Algorithm Design

Abstract

Genetic algorithms are traditionally formulated as search procedures that make use of selection, crossover and mutation operators to implement the search process. However, in recent times, there has been a growing interest in the GA community to replace the traditional two-parent recombination version of genetic algorithms by building and simulating probabilistic graphical models as the core decision making framework. In this new approach, the models guide the exploration of the search space by constructing the distribution of promising solutions and subsequent forward sampling from the distribution at every evolution step until convergence. In this paper, we survey the current literature of research towards this direction, and also give a detailed exposition of one variant of probabilistic graphical model, namely Bayesian network, which arguably subsumes and generalizes many other models mentioned in the literature.

Indexing terms:

• Genetic algorithms
• Bayesian network
• Multinomial distribution
• Model selection
• Parameter estimation