Estudo do conceito de serendipidade como base para novas abordagens ao problema da convergência prematura

Abstract

In the literature, it is common to find many engineering problems which are used to present the effectiveness of the optimization algorithms. Several methods of Bio-Inspired Computing have been proposed as a solution in different contexts of engineering problems. Among these methods, there is a class of algorithms known as Swarm Intelligence. Despite the relative success, most of these algorithms faces a common problem known as premature convergence. It occurs when a swarm loses its ability to generate diversity and consequently converges to a suboptimal solution prematurely. There are several approaches proposed to solve this problem. This doctoral thesis proposes a new approach based on a concept called serendipity. It is usually applied in the field of Recommender Systems. To validate the feasibility of adapting this concept to the new context, a variant called Serendipity-Based Particle Swarm Optimization (SBPSO) has been implemented considering two dimensions of serendipity: chance and sagacity. To evaluate the presented proposal, two sets of computer experiments were performed. Sixteen reference functions which are common in the evaluation of optimization algorithms were used. In the first set of experiments, four functions were used to compare SBPSO to Particle Swarmoptimization (PSO) and some literature variants. In the second ones, twelve other functions were used, but for high dimensionality and a larger number of evaluations of the objective function. In all experiments, the results of the SBPSO were promising and presented a good convergence behaviour with regard to: a) quality of the solution, b) ability to find the global optimum, c) stability of solutions and d) ability to resume the swarmmovement after stagnation has been detected.