Cooperative Multi-Agent Ensembles for Multi-Objective Optimization

Multi-objective optimization (MOO) using metaheuristics is a hot research area that is receiving great interest of algorithm designers. In literature, use of multiple metaheuristic algorithms within a multi-agent framework or within an ensemble system were studied and significant performance improvements compared to single algorithm’s successes is achieved. The first proposed architecture in this thesis study is a multi-metric and multi-deme multi-agent system which comprises several MOO method agents namely: MOGA, SPEA2, MODE, MOSA and MOPSO. The agents cooperate in consecutive sessions to discover and extract a feasible and high-quality Pareto fronts. The system divides the population to sub-populations and assigns them to metaheuristic agents in the beginning of sessions. Once after running metaheuristics, they return the optimized sub-populations to be used in next session and to update the global archive. Four metrics are used within the system in which three of them are multi-objective assessment metrics and one of them is metaheuristic performance measurement metric. The performance of metaheuristics is used by system to adjust their associated number of fitness evaluations. Also, it is used to accept or reject the improved sub-populations. The sub-populations are mixed to get the common population to be used in next session. Meanwhile, the nondominated solutions of each sub-population are used to form the global archive. The global archive keeps all Non-Dominated Solutions discovered by all metaheuristics. In the second architecture a dynamic metaheuristic network is proposed based on a layered platform. The network represents the MOO algorithms with nodes and flow of sub-populations by edges. The system operates in consecutive epochs in which each epoch begins with assigning sub-population to nodes continues with running each metaheuristic within its algorithmic framework. Afterwards the enhanced subpopulations are transferred to all forward linked nodes. At the end of each epoch the metaheuristic agents at layers are changed by a rotation operator. The proposed method contains seven different MOO metaheuristic algorithms formed as 3-3-1 network configuration with three layers. Enhanced sub-populations are transferred to the next layer nodes (one or more layers) when an epoch or session terminates. At the end of all sessions, all local and global Non-dominated individuals are merged. The evaluation of the proposed methods are carried out using a set of well-known benchmarks and the obtained results are compared to state-of-the-art methods. Likewise, it is noticed that the proposed methods outperform the existent state-ofthe-arts methods in the majority of test problems.