Towards memoryless model building

Probabilistic model building methods can render difficult problems feasible by identifying and exploiting dependencies. They build a probabilistic model from the statistical properties of multiple samples (population) scattered in the search space and generate offspring according to this model. The memory requirements of these methods grow along with the problem size as the population must be large enough to guarantee proper initial-supply, decision-making and accurate model-building. The paper presents a novel model based trajectory method, which samples only one point at the time and infers the problem structure online by means of Artificial Neural Network based machine learning technique. As case study we show how the proposed method can very efficiently address hard, non-separable building-block problems, specially designed to be solvable only by population based recombinative methods. The small memory requirement and fast convergence of the proposed method comes at the cost of ...

• Type: Conference
• Year: 2008
• Where: GECCO
• Authors: David Iclanzan, Dumitru Dumitrescu
• Keywords: GECCO 2008, Model Building, Optimization, Probabilistic Model, Probabilistic Model Building