Genetic Algorithm for the solution of majority problem

• Autor(a) principal: Hygor Piaget Monteiro Melo
• Data de Publicação: 2011
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFC
• Texto Completo: http://www.teses.ufc.br/tde_busca/arquivo.php?codArquivo=7145
• topic: FISICA ESTATISTICA E TERMODINAMICA

Many natural and social systems exhibit globally organized behavior without the aid of a centralized control. Examples of such decentralized systems include conventions and norms, social learning in animals and humans, as well as fads, rumors and revolts. Examples are also abundant in biology: the evasive behavior of animals in large groups, such as fish and birds, show a great synchronicity even in the absence of an leader. In order to understand these decentralized systems, one must first understand strategies for global coordination that use only local information. This work explores the use of Genetic Algorithms in the creation of naturally efficient strategies in noisy environments. Genetic Algorithms are an important new tool in problem solving, and offer insight into how evolution may work. By using what is known about genetic algorithms, one can discover more about evolution and its mechanisms. The density classification task is used here to test strategy success, and revealed to be a good test for system-wide coordination and global information processing. Since it is very difficult to evolve highly fit rules when the number of neighbors $k$ is greater than 5, this suggests that evolution may build complex solutions based on solutions to simpler problems. Using this idea, we propose a method to promote rules increasing $k$. Based on the evolution of initial rules with few neighbors and using noise as evolutionary pressure, we were able to find efficient rules for a large number of neighbors, under the condition of a very high noise level. We find that the evolved rules are more robust to noisy environment than the majority rule. This increased efficiency at higher noise levels can be explained in terms of the larger weight given by these rules to the information of the evolving agent itself (not influenced by noise) than to the information obtained from its neighbors. As a consequence, the agents using these evolved rules tend to keep their own states, unless the great majority of their neighbors disagree with them, showing a persistence behavior that can be seen in social experiments.