Research On Evolutionary Computationmethod With Blind Evaluation

As an important branch of artificial intelligence,evolutionary computation is an effective way to solve complex optimization problems.Individual fitness evaluation is a key step in evolutionary computation.The integration,efficiency,and accuracy of individual fitness evaluation methods have major influence on the feasibility,practicability,and accuracy of evolutionary computation.Due to the lack of integrated and efficient individual fitness evaluation method,evolutionary computation can not effectively solve optimization problems.This thesis studies the “evolutionary computation method with blind evaluation”.By introducing a blind evaluation method of individual fitness and designing an associated evolutionary computation method,the performance of evolutionary computation can be improved by a large margin,and its application scope can also be expanded.The main contributions of this thesis are summarized as follows:1.The blind evaluation method of individual fitness is studied.First,based on the characteristics of optimization problems and evolutionary computation,we propose a concept of the blind evaluation of individual fitness,and expound a definition of the individual fitness blind evaluation method.Second,based on convolutional neural networks,we take the blind evaluation neural network as a specific form of blind evaluation method.Finally,we analyze the operational complexity and running efficiency of the blind evaluation neural network.The evaluation accuracy of the blind evaluation neural network is verified by experiments.2.The proposed blind evaluation method is applied to the evolutionary computation.First,the structure of the blind evaluation neural network is optimized to improve the performance of evolutionary computation.At the same time,two advanced strategies are proposed to design the evolutionary computation method with blind evaluation.The two advanced strategies include the principle of “reducing the number of evaluations and increasing the number of evaluated individuals” and the strategy of “outputting staged best individuals”.The former one is used to improve the efficiency of evolutionary computation,while the later one is used to improve the accuracy of evolutionary computation.Second,the blind evaluation method is combined with the shuffled frog leaping algorithm,and an evolutionary computation method with blind evaluation,SSFLA-BE,is proposed.Finally,the efficiency and accuracy of the SSFLA-BE are verified by experiments.3.In order to further improve the performance of the evolutionary computation method with blind evaluation,the blind evaluation method is further enhanced.This thesis proposes the “regional reinforcement learning” strategy to improve the accuracy of the evolutionary computation method with blind evaluation,and proposes the “transfer learning” strategy to improve the feasibility and practicability of the evolutionary computation method with blind evaluation.The experimental results show that the proposed strategies effectively improve the performance of the evolutionary computation method with blind evaluation.4.From the perspective of optimization,solving local dimming problems with evolutionary computation is studied.First,to improve the displayed image quality of local dimming system,an integrated and rough individual fitness evaluation method with high evaluation accuracy is designed.Then the fireworks algorithm is adopted in combination with the proposed evaluation method to solve local dimming problems.Then a local dimming algorithm based on evolutionary computation,FWA-LD,is proposed.Finally,The searching ability of the FWA-LD is verified by experiments,and the high accuracy of the proposed evaluation method is also proved by subjective tests.5.In order to obtain a more efficient local dimming method,the evolutionary computation method with blind evaluation is applied to local dimming problems.By solving local dimming problems,we further verify the performance of the evolutionary computation with blind evaluation.First,we propose a structure and a training strategy for the blind evaluation neural network which is used to solve local dimming problems.Second,the SSFLA-BE and the FWA-LD are fused to generate a new local dimming method,namely SSFLA-FWA-LD.Finally,experimental results have shown that the proposed local dimming method achieves higher efficiency.