Search On Evolutionary Dynamic Optimization Algorithms And Applications

Dynamic optimization problems (DOPs), whose objectives and constraints change with the time, are common in real world. The main concern of the tra-ditional optimization community has been static optimization problems. Though many successful methods have been developed for static optimization, but these methods are not suitable for DOPs. Being population-based and self-adaptive, Evolutionary Algorithms (EAs), however, have instinct advantage in solving DOP-s. Up to now, applying EAs to DOPs has turned out to be a current research focus in optimization and intelligent computation communities. This dissertation is de-voted to memory-based EAs for DOPs. The main idea of the memory-based EAs is to store information from the past, and reuse them to enhance the performance of EAs for DOPs in the future.The concerns of this dissertation include the memory updating strategies, combining memory schemes with multipopulation algorithms, the organization structures of memory and the application of memory-based EAs. The main con-tributions of the dissertation are as follows.(1) Improving the classic similar strategy for memory updating. In order to store as more as possible information in the memory of limited capacity and main-tain the diversity of memory individuals, the classic similar strategy replaces the memory individual that is closest to the new individual. But several illustrations given in this dissertation demonstrate that the similar strategy fails to maintain good memory diversity in some cases, and that keeping memory individuals that are used frequently results good performance. Based on the illustrations, a adap-tive updating strategy is proposed. This strategy analyzes the memory updating behavior and records the using frequencies of the memory individuals. When up-dating the memory, it protects the individuals of high using frequency from being replaced, and eliminates those have not been used for a long time.(2)The combination of memory and multipopulation algorithms. Multipop-ulation algorithms are an important kind of EAs for multimodal DOPs. Almost all the existing memory schemes were proposed with singe population algorithm- s, and simply applying them to multipopulation algorithms does not necessarily result in positive effect. The memory scheme proposed in this dissertation, uses a fitness-independent updating strategy, consequently, it is able to enlarge the memory capacity at an acceptable degree of computational overhead. When up-dating the population with memory, both the fitness and the space distribution of the memory individuals are considered. Therefore, the disturbance to the search of multipopulation algorithms from the memory scheme is relieved. The experi-mental results show that the tested typical memory schemes sometimes harm the performance of the multipopulation algorithms, but the enhancing effect of the proposed memory scheme is significant.(3)Employing binary space partition tree as the memory structure. Memory structure is important, which determines the operations of the memory scheme and the computational complexity. However, up to now, there is no special study on this subject. All the existing memory schemes use linear list as the structure. This dissertation proposes a neighbor-shift memory tree scheme that uses the binary space partition tree as the structure. The scheme partitions the search space into subregions. Each regions contains a memory individual. When the search of EAs reaches a subregion for the first time since the last environmental change, the memory individuals of that subregion will be retrieved. Compared with the schemes using linear list, this schemes possesses several good features such as unlimit memory capacity, parameterless and adaptively recording more information from the regions where the good solutions appear frequently.(4)According to the properties of the dynamic optimal reactive power prob-lem, this dissertation proposes a memory-enhanced differential evolution. This algorithm records the environmental attributions of the power grid and the cor-responding dispatch solutions. Two strategies are proposed to retrieve memory information after the grid changes. The first one stochastically select memory dispatch solutions from the memory into the population of differential evolution according to the matching degree of the current and the memory environmen-tal attributions. The second strategy makes use of the statistic properties of all the memory entries to generate immigrants for the population. Experiments on IEEE30-bus system and IEEE118-bus system demonstrate that the proposed memory scheme can dramatically enhance the performance of the differential evo-lution to solve dynamic optimal reactive power.