| On 2018-01-09 at 11:00:00 (Brussels Time) |
Abstract
In general, meta-heuristics identify promising regions of the search space and exploit them to obtain the best quality solutions from those regions. However, meta-heuristics based approaches present typical difficulties: getting trapped in local optima, have convergence problems, among others. For this reason, the design of strategies for improving the search of metaheuristics has been severally studied. In this presentation we will talk about different learning strategies focused on detecting Quasi-near-optimal (Q-n-o) solutions, candidate solutions that appear promising but end up to not be of good quality solutions. For this, a division of the search process of a metaheuristic algorithm has been proposed: a first step for learning about Q-n-o solution for a defined problem and, a second step for solving it using this knowledge. These strategies are inspired by Opposition-Based Learning literature and the key idea is to influence the search process of a defined metaheuristic to reach more interesting region s of the search space. Also, an evaluation of three strategies will be presented, applied to two well-known ant algorithms: Ant Solver for Constraint Satisfaction Problems and, Ant Knapsack for solving the Multidimensional Knapsack Problem. Moreover, some new research related to an initialization procedure for Population-Based meta-heuristics will be presented, using information of Q-n-o candidate solutions.
Keywords
metaheuristics, ant colony optimization, opposition-based learning
