| dc.description | This thesis presents a new formulation for the swarm optimization technique as a system of autonomous agents. The proposed technique is based on an intimate understanding of swarms and animal aggregations in an attempt to simulate cognitive thinking of their members. The dynamic balance between gregarious and social intolerance behaviors demonstrated by social animals is used to form the swarm and keep its persistence. In this work, members of the swarm are represented by agents that enjoy a certain degree of freewill to respond adaptively to changes on the states of their swarm mates. Adaptive responses are reflected on the way agents move inside the problem domain. A new set of basic behaviors is defined, namely imitation, memory retrieval, momentum, and play. A multi-criterion decision making process (MCDM) is employed to update positions of swarm members in the problem space. Decision making alternatives are defined from the set of basic behaviors. Fitness and diversity characterize the decision criteria that are used to measure the performance of each alternative. Levenshtien edit distance is used to measure the distance between agents in the genotype space. Criteria are then standardized by means of fuzzy sets. Fuzzified values of criteria are aggregated by the fuzzy ordered weighted average (OWA) to reach a single evaluation function. The overall decision making process is made to promote both fitness and diversity. The proposed technique is tested using the traveling salesmen (TSP) and quadratic assignment problems (QAP). Results and comparisons show that the technique outperforms the traditional particle swarm optimizer (PSO). Also, a comparison of the proposed technique with the standard genetic algorithm (SGA) shows that comparable results can be obtained. An extension of the proposed technique is also proposed to solve optimization problems with continuous variables. For this class of optimization, a large set of diverse benchmark problems is used to test the proposed technique. A comparison of the performance with the simple evolutionary algorithm SEA and many other particle swami variants is also carried out. Results show that the proposed technique outperforms other techniques included in the comparison in almost all the tested problems. | en_US |
