Towards Efficient Multi-objective Optimization Based On Knowledge Transfer

Multi-objective optimization problems(MOPs)are widely existed in the real world,which refer to the optimization problems involving several conflicting objectives that need to be optimized simultaneously.For example,scheduling in logistics is essentially a MOP,which needs to balance both the service cost and customer satisfaction during the decision making.Multi-objective optimization aims to find a set of Pareto optimal solutions,rather than a single solution in the context of single-objective optimization.For solving MOPs,compared with traditional mathematical methods,population based evolutionary algorithms(EAs)has been recognized as a powerful tool,which possess strong search capability and are easy to implement.In recent years,to improve the efficacy of multi-objective evolutionary algorithms,many works focus on designing efficient search paradigm with knowledge transferred across related problems.However,how to find related problems with helpful optimization knowledge is a non-trivial task.For MOP,it contains several single optimization objectives,indicating a nature relevance between the single-and multi-objective problems.If we can learn and transfer the knowledge from these single objective problems(SOPs),it could improve the efficiency of multi-objective optimization significantly.Taking the cue,focusing on how to leverage the SOPs and to conduct knowledge transfer for efficient multi-objective optimization,this thesis carries out the study of evolutionary multi-objective optimization based on knowledge transfer from SOPs.Specifically,the main contributions of this thesis are summarized as follows:First of all,we propose a new algorithm of multi-objective optimization based on knowledge transfer from single-objective problems,called MOST.The algorithm configures the decomposed SOPs as the sources for transferred knowledge.A single-layer denoising autoencoder is then employed for knowledge transfer.Experimental results on the common benchmarks show that the MOST can speed up the convergence without loss in solution quality.Further,deeper analysis shows that transfer success rate has a positive correlation with the efficacy of the proposed paradigm,indicating that it could be used as the quality indicator for future researches.Next,to reduce the negative effects caused by MOST,we propose an improved algorithm for multi-objective optimization based on integrated knowledge transfer from SOPs obtained by weighted-sum decomposition,termed MOIST.Moreover,decomposition methods for constructing SOPs and knowledge transfer method are also explored.In the MOIST,the SOPs are designed by weighted-sum method,while knowledge transfer is based on an integrated structure.Empirical studies show that the proposed method can reduce negative knowledge transfer and improve the convergence speed of multi-objective optimization.In addition,preferences on problems with distinct characteristic in terms of different knowledge transfer and single-objective constructing methods are also analyzed,providing some inspirations for future works.