| With the rise and development of smart cities,many optimization problems with complex characteristics are emerging in areas such as intelligent traffic scheduling and smart logistics scheduling.These problems often possess the characteristics of being computationally expensive,time-consuming,and requiring multi-task optimization.Traditional evolutionary computation(EC)methods are ineffective in dealing with these complex optimization problems and suffer from performance degradation.Although existing research has proposed some methods to incorporate problem domain knowledge to assist EC for better optimization,most of them are single-source knowledge incorporation methods.That is,existing single-source knowledge incorporation-based EC methods are limited to using single-form knowledge within the problem domain.However,the complementarity of different problem-and algorithmrelated multi-source knowledge in single problem domain can be utilized to improve search efficiency.Moreover,the similar and common knowledge of different problem-and algorithmrelated multi-source knowledge in multiple problem domains can also be utilized to improve search efficiency.Therefore,in view of the performance bottleneck of existing EC methods in solving complex expensive and multitask optimization problems,this thesis conducts research on multisource knowledge learning-based EC.This thesis focuses on three key aspects of the multisource knowledge learning process,including knowledge mining,knowledge selection,and knowledge transfer.Specifically,this thesis proposes a multi-level smoothness regularizationbased EC method for knowledge mining,proposes a multi-scale fitness evaluation-based EC method for knowledge selection,a multi-population evolution knowledge transfer-based EC method,and a cross-task mapping-based knowledge transfer-based EC method for knowledge transfer.The proposed methods can help achieve a better balance between computational time and solution accuracy,and improve the search efficiency and global search ability of EC methods.The proposed methods are applied in multiple fields of smart cities to validate their effectiveness.The main research work and contributions of this thesis include:(1)A multi-level regularization knowledge mining-based EC method is proposed.The proposed method can mine effective and generalizable multi-source knowledge from historical data collected in the real-world application process.To overcome the challenges of dynamical changes in environments and computational expensiveness in the real-world application problem,this thesis treats multiple mapping models as different knowledge and proposes a learning to adapt(LA)framework to mine this knowledge.The mapping model obtained by the LA framework can quickly and efficiently adapt to new environments.For the LA framework,a multisource transfer-based particle swarm optimization algorithm is proposed to construct a high-quality training dataset,which can effectively alleviate the problem of under-optimizing.Based on the obtained dataset,a multi-level smoothness regularization knowledge mining method is proposed to learn multiple mapping models as multi-source knowledge,which can effectively alleviate the problem of overfitting in the learning process and obtain mapping models with good generalization performance.The proposed method is applied to the traffic signal timing optimization problem in field of intelligent transportation,and the experimental results validate the effectiveness of the proposed knowledge mining method.(2)A multi-scale fitness evaluation knowledge selection-based EC method is proposed.Considering the formulated multi-source knowledge with different accuracy scales and computational costs in the problem domain,this thesis treats different fitness evaluation methods as different knowledge,and proposes a scale-adaptive fitness evaluation(SAFE)method to adaptively select fitness evaluation knowledge to improve the optimization performance on expensive optimization problems.In the proposed SAFE method,a one-way switch strategy and a two-way switch strategy are proposed for the switch between different evaluation methods.The selection of evaluation methods is based on population evolution status and search requirements,which is reflected by the proposed contribution factors.Additionally,a fast neighbor best-based evaluation strategy is proposed to fully exploit the advantages of the SAFE method.The proposed method is applied to the crowdsourcing logistics scheduling problem in the field of intelligent logistics,and the experimental results validate the effectiveness of the proposed knowledge selection method.(3)A multi-population evolution knowledge transfer-based EC method is proposed.The proposed method accurately identifies and utilizes similar knowledge across multiple populations to improve search efficiency for multitask optimization problems.To overcome the shortcomings of traditional EC methods such as insufficient knowledge utilization and low efficiency in solving multi-task optimization problems,this thesis treats different algorithm parameters used on different tasks as different knowledge,and proposes to transfer this knowledge between similar tasks.A task representation strategy and a task grouping strategy are proposed to accurately identify and group relevant and similar tasks.Based on the selected relevant knowledge,a novel successful evolution experience-based knowledge transfer method is proposed,which uses the algorithm parameters of similar source tasks to assist efficient evolution on the target task,thereby improving the overall optimization performance on multiple tasks.The proposed method is applied to a multi-task robotic arm control problem in intelligent manufacturing,and the experimental results validate the effectiveness of the proposed knowledge transfer method.(4)A cross-task mapping-enabled knowledge transfer-based EC method is proposed.To overcome the challenges posed by knowledge heterogeneity and to transfer and utilize similar source knowledge to improve search efficiency,this thesis treats different population distribution for different tasks as different knowledge,and proposes to perform cross-task mapping between tasks to make knowledge transfer between heterogeneous tasks viable.To reduce negative transfer caused by the differences in the importance of dimensions between tasks,an orthogonal transfer method is proposed to efficiently utilize source tasks’ dimensional information.Moreover,a cross-dimensional transfer method is proposed to capture and utilize cross-dimensional similarity to further improve transfer quality.The proposed method is applied to the multi-task robotic double pole balancing problem in intelligent manufacturing,and the experimental results validate the effectiveness of the proposed knowledge transfer method.In summary,this thesis conducts research on the multi-source knowledge learning-based EC method.Through sufficient experimental studies and analysis,this thesis verifies the effectiveness of the proposed methods and applies them to solve complex application problems in smart cities.This research not only provides new efficient algorithms for complex optimization problems,including expensive optimization and multi-task optimization,but also promotes the development of EC in solving complex optimization problems. |
