Ant Colony Optimization in Stationary and Dynamic Environments
thesisposted on 14.06.2013, 09:45 by Michalis Mavrovouniotis
The ant colony optimization (ACO) metaheuristic is inspired by the foraging behaviour of real ant colonies. Similarly with other metaheuristics, ACO suffers from stagnation behaviour, where all ants construct the same solution from early stages. In result, the solution quality may be degraded because the population may get trapped on local optima. In this thesis, we propose a novel approach, called direct communication (DC) scheme, that helps ACO algorithms to escape from a local optimum if they get trapped. The experimental results on two routing problems showed that the DC scheme is effective. Usually, researchers are focused on problems in which they have static environment. In the last decade, there is a growing interest to apply nature-inspired metaheuristics in optimization problems with dynamic environments. Usually, dynamic optimization problems (DOPs) are addressed using evolutionary algorithms. In this thesis, we apply several novel ACO algorithms in two routing DOPs. The proposed ACO algorithms are integrated with immigrants schemes in which immigrant ants are generated, either randomly or with the use of knowledge from previous environment(s), and replace other ants in the current population. The experimental results showed that each proposed algorithm performs better in different dynamic cases, and that they have better performance than other peer ACO algorithms in general. The existing benchmark generators for DOPs are developed for binary-encoded combinatorial problems. Since routing problems are usually permutation-encoded combinatorial problems, the dynamic environments used in the experiments are generated using a novel benchmark generator that converts a static problem instance to a dynamic one. The specific dynamic benchmark generator changes the fitness landscape of the problem, which causes the optimum to change in every environmental change. Furthermore in this thesis, another benchmark generator is proposed which moves the population to another location in the fitness landscape, instead of modifying it. In this way, the optimum is known and one can see how close to the optimum an algorithm performs during the environmental changes.