Some Theoretical Studies Of Elman Neural Networks And Evolutionary Algorithms And Their Applications

Computation Intelligence (CI) is the theory and method of information obtaining,processing and applying by simulating human's intelligent mechanism or living'sevolution processing or human (or other species)'s intelligent behavior on moderncomputing tools. Computation Intelligence has three main fields, namely that: ArtificialNeural Networks (ANN), Fuzzy System (FS) and Evolutionary Algorithms (EA),respectively. This thesis has performed some theoretical and application studies in thefields of ANN and EA. Of ANN, aiming at the Elman NN which includes the partfeedback linkages, the paper has studied its algorithm, convergence and applicationresearch. Of EA, the algorithm and application research of Genetic Algorithm (GA),Particle Swarm Optimization (PSO) and hybrid evolution algorithms are performed. Themajor contents could be summarized as follows:I The theoretical studies of Elman Neural Networks (Elman NNs) and theirapplications for ultrasonic motor (USM)(1) Considered the back propagations of the output-layer nodes, the authordeveloped two improved models based on Elman NNs, namely that the Output-InputFeedback Elman (OIF Elman) Neural Networks and the Output-Hidden Feedback Elman(OHF Elman) Neural Networks. The two models have added the linkage fromOutput-layer to Input-layer and from Output-layer to Hidden-layer, respectively. Theirlearning algorithms are derived according to gradient descent method. In order toguarantee the convergence of the models, the adaptive learning rates of neural networksare derived using discrete-type Lyapunov stability analysis. As a application, the ElmanNNs and the improved models are applied to the speed identification of ultrasonic motors.The result comparisons and analysis are also performed.(2) Aiming at the control of USM, the paper developed a Recurrent BackPropagation Controller (RBPC). Our proposed OIF Elman model is applied in thecontroller as an identifier for USM with the usage of providing a passage for error's backpropagation. The learning algorithm is derived by like gradient descent method also.Furthermore, in order to guarantee the stability,the convergent conditions are providedaccording to the Lyapunov stability theorem. Simulations are executed and thecomparisons to those of reference method are performed also. The results show that theproposed method has good performances.II Some algorithmic studies of evolutionary algorithms and their applications(1) By introducing the migration strategy which called intermarry-strategy intomultiple-population genetic algorithm, the author proposes an improved geneticalgorithm, namely that the Tribe-Intermarry-Strategy Genetic Algorithm (TISGA). In themethod, after every given number iterations, a mount of number individuals are selectedrandomly from different populations according to a certain principle. Then the selectedindividuals reproduce the same number of offspring, which should be return to theoriginal populations. The numerical results show the effectiveness of the improved GA.(2) Focused on the selection operator of GA, the author develops another improvedGA. In selection process, the individuals should be selected pair by pair. The first one of apair is selected according to the candidates' fitness, but the second one not only accordingto the fitness, but also according to the distances from the candidates to the selected one.By taking the two factors into account, the author gives a novel selection operator.Therefore the improved method is called by improved genetic algorithm based onFitness-Distance selection mechanism. Simulations are also performed to verify theeffectiveness of the proposed method.(3) According to the natural features, especially, the population-size-changing ruleon human beings, the author presents a novel improved GA with variable population-size(VPGA) which obtains a better balance between the variety and continuity. There aremainly two new points are taking into account by VPGA. First, of most feature species,parents are neither dead after their reproduction right away, nor living forever. In fact, theindividual is apt to die when it is old. In VPGA each individual is designated a dyingprobability according to its living generations. Second, the size of features' populationalways increases, until there is a contagious disease or a war breaking out among thepopulation. In general, the disease or the war will reduce the size of the populationsharply. In the VPGA, we reduce the size of the population to the initial size when itreaches the given limitation, those individuals with higher fitness have moreopportunities to survive.(4) Illustrated by "swap operator", the "subtraction" operator between two particlepositions of PSO is modified. Therefore a discrete PSO method is constructed for the TSP.The crossover eliminated technique is also executed in the proposed method. Numericalresults show that the proposed method could improve the size of resolvable TSPproblems.(5) Based on the "generalized chromosome" coding technique of the GTSP, theabove proposed method is extended for solving the GTSP. Two local search techniquesare also added into the method. 19 GTSP problems are examined for benchmarking. Theresults show that the method is effective for solving the GTSP problem. To the best of ourknowledge, there has been no attempt in proposing a PSO-based algorithm for the GTSPso far. The proposed PSO-based algorithm could provide a suitable approach for solvingthe GTSP.(6) In the paper the hybrid evolutionary algorithms are summarized into threepatterns, namely that parallel-hybrid pattern, series-hybrid pattern and idea-hybrid pattern.Taking GA, PSO and IA for instances, different hybrid patterns are described concretely.Some benchmark functions are also examined for comparison and analysis.To sum up, the paper develops two improved Elman NNs, the learning algorithmand stability conditions are also studied. The work enriches the researches of Elman NNs.Especially the application of Elman NNs to the identification and control for USM speednot only expands Elman NNs' application fields, but also presents a novel idea for USMspeed identification and control. Three improved GAs are developed from differentaspects, abundant numerical results are also detailed and analyzed. The work enriches thealgorithmic research of GA in a certain extend. A discrete PSO method for TSP isproposed, it could improve the size of resolvable TSP problems greatly compared withthe existed PSO-based method. Furthermore, the method is extended for solving theGTSP. To the best of our knowledge, there has been no attempt in proposing a PSO-basedalgorithm for the GTSP so far. In the end, the hybrid evolutionary algorithms aresummarized into three patterns, which are described concretely by taking GA, PSO andIA for instance.