Research On Dynamical Resource Allocation Of An External Archive Guided Multiobjective Evolutionary Algorithm

The goal of multiobjective optimization is to approximate a set of Pareto optimal solutions,which represents the trade-off relations among different objectives. Evolutionary computation(EC),due to its population-based nature, has been widely used in addressing multiobjective optimization problems(MOPs). More recently, decomposition-based mulitiobjective evolutionary algorithm(MOEA/D) has attracted a wide range of attentions. MOEA/D decomposes a MOP into several single objective subproblems and optimize them simultaneously. In the original MOEA/D, the computational resources are equally allocated to all subproblems. However, a strategy for dynamic resource allocation becomes very necessary to further improve the efficiency of the algorithm, as some subproblems may be more difficult to solve than others. This thesis mainly investigates different strategies for dynamic resource allocation, based on the guidance of an external archive under the decomposition-based MOEA framework. The main contributions of this thesis are as follows.First, a hybrid MOEA framework is adopted where both a decomposition-based working population and a Pareto-based external archive are adopted. A two-phase strategy for dynamic resource allocation is proposed. The evolutionary process are explicitly divided into two phases. First,the convergence information from the external archive is utilized to guide the working population.And the diversity information is used in the second phase.Second, for some special optimization problems where the evolutionary process can not be explicitly divided into the two phases, a multi-phase strategy for dynamic resource allocation is proposed. Based the evolutionary status(convergence or diversity), a switching mechanism is adopted to adaptively use either convergence or diversity information in the external archive, to guide the evolutionary search in the working population. The balance of the convergence and diversity is improved by this strategy.Finally, the experiments on MONRP and MOTSP show that the proposed algorithms, allocating computing resources reasonably, outperformed other state-of-the-art algorithms. And the effect of the dynamic resource allocation is proved.