| Genetic Algotithm is a kind of highly parallel adaptive random search method which is advanced by professor. Hollad originally. It is based on the biological nature selection and evolutionism together. In recent years people are continuing to improve and develope it, applying it into such complex problems which couldn't be solved in the traditional methods, e.g. advantageous making-up combinatorial optimization, image processing and numerical value optimization. Now it has become a heated topic.Genetic Algorithm uses the simple codes to express different complex structures of questions, the selection, crossover and mutation operators to the population are not dependent on the question, while denning the searhing direction according to the natural selecting rule of survical of the fittest simply. This is a sort of directed random selection. Therefore this way is suitable to large par-ellel dealing. It is unbound of the condition of searching space differentiability, single-peak, continuity and there is no need of other assistant instrument. These features not only make the genetic algorithm become higher efficient and easy to operate, but aslo make it possess global optimality., implicit parallelism, robustness and general. This method also has some defects, e.g. the slow speed of convergence, the badness of stability and operation, premature convergence.Genetics Based Machine Learning is one important aspect of current ge-netic algorithm research. The most outstanding study is the research of classifier system. The genetic machine learning system based on classifier system is composed of credit assignment algorithm, the rule and message system and genetic algorithm. Holland presented the first classifier system which is centered on the genetic and the bucket brigade algorithm in 1986.This article combines the ideas between genetic algorithm and machine learning, gives some important improvements on the classifier system and puts forward the improved genetic algorithm based on machine learning.(1) The application of strengthening factor.In the credit assignment, the reward to the winning classifier ensures the existence of the best individual and strengthens local searching ability of this algorithm, making the population approach the most excellent solution continuously.(2) The application of crowding factor.We use the crowding factor both in the rule and message system and the process of the genetic algorithm. We replace the worst classifiers with the best conditional messages after each machine learning; in addition, we take the place of the worst individual in the original group with the most similar better one of the filial generation produced by the crossover operator after each genetic algorithm.The introduction of crowding factor solves the contradiction between the selection pressure and the population diversity. It not only ensures the existence of the best individual, but also keeps the population diversity.(3) The application of combine factor.We combine similiar classifiers after each algorithm perfoment and make the average value of all similiar classifiers as the final strength value. This way prevents the appearance of super individual and avoids premature convergence be-cause of the gradual narrowing of searching strap, maintaining the former searching space.(4) The calculation of the credit assignment in improved genetic algorithm based on machine learning.Let the strengh of classifier C in time t be S(C, t) , the bid coefficient be Cbid, the random noise in the valid bid be N(bid), the bid tax coefficient be Cbidtax, the life tax coefficient be Clifetax, the bid control parameter be1, when it bidthe old winner be m, the new one be m + 1,0, when it dosen't bid , the reward to the winner be R', the repay be R(t), and R(t) = Bid(m,t).When C becomeing the candidate, the bid value is The valid bid isEBid = Bid + N( bid)The tax isWhen the candidate classifier C takes part in biding a message, its strengh isStrength(C, t + 1) = Strength(C, t) - Bid(C,... |
PDF Full Text Request