Aplicações de mecânica estatística a especiação simpátrica e inferência aproximativa / Applications of statistical mechanics to sympatric speciation and aproximative inference

AUTOR(ES)
DATA DE PUBLICAÇÃO

2009

RESUMO

This thesis presents applications of the framework of Statistical Mechanics to two independent problems. The first corresponds to a computational model for the evolution of Assortative Mating in the Sympatric Speciation process; and the second a learning algorithm built by means of a Bayesian Inference approach. In the biological problem each individual in an agent-based model is composed of two traits. One trait, called the ecological trait, is directly related with the fitness; the other, called the marker trait, has no bearing on the fitness. The traits are determined by different loci which are linked by a recombination rate. There is also the possibility of evolution of mating preferences, which are inherited from the mother and subject to random variations. The study of the phase diagram in the spa e of parameters describing the environment (like carrying capacity and disruptive selection) reveals the existence of three phases: (i) assortative mating; (ii) extinction of one allele from ecological loci; and (iii) Hardy-Weinberg equilibrium. It was verifed that the assortative mating an emerge or even be lost (and vice-versa) acording with the environmental hanges. Moreover, the system shows memory of the initial condition, characterising a hysteresis. Hysteresis is the signature of first order phase transition, which allows the description of the system by means of the Statistical Mechanics framework. In relation to the Bayesian Inference, a supervised learning algorithm was constructed by means of the Expectation Propagation approach. The idea is to estimate the parameters which compose a Teacher Perceptron by the substitution of the original posterior distribution, intra table, by a tractable approximative distribution. The step-by-step update of the terms composing the approximative distribution was performed by using the Expectation Propagation algorithm. The update must be repeated until the convergence ocurrs. Using the Central Limit Theorem and the Cavity Approah, it was possible to get a generic algorithm that has shown a very good performance in two application scenarios: The Binary Perceptron Model and the Gaussian Perceptron Model.

ASSUNTO(S)

bayesian inference especiação simpátrica sympatric speciation modelos de mecânica estatística assortative mating inferência bayesiana statistical mechanics models

ACESSO AO ARTIGO

Documentos Relacionados