GAGA: Model Building Genetic Algorithms Using Sub-population and Sub-probability Vector

The Compact Genetic Algorithm (cGA) has a distinct characteristic that it requires almost minimal memory to store candidate solutions. The probability vector is used to generate candidate solutions. This probability vector represents a structure of the population as a probability distribution over the set of solution. It has been established that the power of cGA is comparable to the standard Simple Genetic Algorithm (sGA) with uniform crossover. Hence, its limitation hinges on the assumption of the independency between each individual bit. For example, a standard difficult test problem for GA is a deceptive function, or so called Trap function. cGA fails to solve this problem. This work proposes another approach of using the probability vector as a model of the structure of solutions, named GAGA (for GA-in-GA). GAGA employs two new methods for updating the probability vector. The first-method uses a sub-genetic algorithm and the second method uses a sub-probability vector. The experimental results show that the proposed methods can solve the problem that has the dependency between bits such as Trap function.