Research Into Multi-Level Multi-Scale Knowledge-Assisted Evolutionary Computation And Its Application

With the booming of intelligence,there is an urgent need for intelligent optimization methods because diverse and complex optimization problems are emerging day by day.Evolutionary computation is a class of general optimization methods inspired by the behavior of evolutionary intelligence and swarm intelligence of nature and organisms,and solves problems optimally by iterative population evolution.Because of its applicability to difficult optimization problems without exact mathematical models,evolutionary computation has been applied to a wide range of optimization problems in many fields with remarkable results.However,traditional evolutionary computation still faces the performance bottlenecks of poor global search,low computational efficiency,and weak multi-solution problem-solving ability when solving emerging complex optimization problems.To address these issues,this paper proposes the idea of multi-level multi-scale knowledge-assisted evolutionary computation for solving optimization problems and conducts research on knowledge-assisted evolutionary computation algorithms at three levels of problem,model,and algorithm,and six scales including two scales of each level,respectively,and applies them to large-scale neural network parameter optimization,arterial traffic signal timing optimization in smart city,multi-robot navigation optimization and logistics vehicle routing problem in epidemics for validation.The main research works and contributions of this paper are listed as follows.First,this paper proposes multi-scale knowledge-assisted evolutionary computation methods at the problem level to enhance the ability of evolutionary computation to solve large-scale optimization problems and expensive optimization problems.The proposed methods include:(1)In the problem-level variable-scale knowledge-assisted aspect,this paper proposes a variable interaction knowledge-assisted large-scale co-evolutionary algorithm.Based on the proposed theory related to the multiplicatively separable function,the proposed algorithm uses the proposed dual differential grouping method to actively learn and extract multiple independent variable interaction knowledge from the optimization problem,so that the problem can be decomposed into several sub-problems by variable interaction knowledge.Then,the co-evolutionary algorithm is used to carry out the cooperative evolution for optimizing sub-problems,which enhances the algorithm’s ability to solve different kinds of large-scale problems globally.The proposed algorithm is validated on well-known competition test problems and a large-scale neural network parameter optimization application problem.(2)In the problem-level function-scale knowledge-assisted aspect,a data-driven evolutionary algorithm assisted by fitness knowledge is proposed.Based on the fitness knowledge,the proposed localized data generation method and boosting strategy are used to build an accurate and efficient data-driven surrogate model,which can replace the real but computationally expensive fitness evaluation,so as to drive the algorithm to solve expensive optimization problems efficiently.This enhances the ability of evolutionary computation algorithms to solve expensive optimization problems globally.The proposed algorithm is tested and validated in widely-used test problems and a smart city arterial traffic signal timing optimization application.Second,this paper proposes multi-scale knowledge-assisted evolutionary computation methods at the model level to improve the parallel efficiency of the algorithm and the efficiency of computational resource usage,thus improving the computational efficiency of the evolutionary computation algorithm.The proposed methods include:(1)In the model-level particle-scale knowledge-assisted aspect,the pipeline-based parallel particle swarm optimization algorithm assisted by the topology knowledge among particles is proposed,which is combined with the idea of pipeline technology to conduct evolution in parallel at the generation level,so that while some particles are performing their contemporary evolution,other particles can enter the next generation or even the generation after next generation to evolve themselves,thus improving the parallel efficiency and the computational efficiency of the algorithm.In addition,the theoretical analysis of the proposed algorithm in term of the acceleration is carried out,and the generation-level parallelism is successfully implemented on different distributed computing platforms,where the results of the time-consuming simulation experiments are consistent with the theoretical analysis.(2)In the model-level population-scale knowledge assistance aspect,the proposed distributed differential evolution algorithm assisted by the topology knowledge among populations is proposed,and the proposed general performance indicator and fitness evaluation allocation method are used to achieve adaptive resource allocation based on topology knowledge among populations,so as to allocate computational resources from poorly-performing populations to well-performing populations,which can reduce the waste of computational resources,improve the overall utilization of resources,and accelerate the algorithm optimization speed.In addition,the proposed algorithm is theoretically analyzed and the theoretical lower bound of its optimization error is discussed,and the algorithm is tested and compared with the state-of-the-art distributed algorithms on time-consuming simulation experiments.Third,this paper proposes algorithm-level multi-scale knowledge-assisted evolutionary computation methods to improve the multi-solution problem-solving ability of evolutionary computation for multi-task and multi-objective problems.The proposed methods include:(1)In the algorithm-level individual-scale knowledge assistance aspect,the individual evolutionary meta-knowledge-assisted multi-task differential evolution algorithm is proposed to improve the cross-domain general problem-solving ability of the algorithm by transferring the meta-knowledge obtained in the individual evolutionary process,so as to promote the multi-solution problem-solving ability of evolutionary computation algorithm for multi-task optimization problems.The proposed algorithm is validated on common competition benchmark problems and multi-robot navigation optimization problems.(2)In the algorithm-level population-scale knowledge assistance aspect,the proposed multi-objective ant colony system algorithm,with the assistance of population evolutionary meta-knowledge,extends the existing multiple populations for multiple objectives framework,uses multiple populations to solve multiple objectives respectively,and generates high-quality solutions that weigh multiple conflicting objectives through the knowledge obtained from the optimization process of multiple populations,so as to improve the multi-solution problem-solving ability of the algorithm for multi-objective problems.The proposed algorithm is validated through the multi-objective vehicle routing problems in epidemics.In summary,this paper carries out the research into the intelligent optimization algorithm based on the idea of multi-level multi-scale knowledge-assisted evolutionary computation,which enhances the global search ability,the computational efficiency,and the multi-solution problem-solving ability of the algorithm,and thus realizes the algorithm in solving complex optimization problems globally and efficiently.The performance improvements of the proposed algorithms have been validated in large-scale neural network parameter optimization,arterial traffic signal timing optimization in smart cities,multi-robot navigation problem,and vehicle routing problems in epidemics.The whole research provides efficient optimization methods for complex optimization problems,contributes to the development of evolutionary computation in the field of complex optimization,and inspires the combination of knowledge-assisted and artificial intelligence methods.