Some Issues On Computational Intelligence For Machine Learning

Machine Learning (ML) is one of the caused branches in Artificial Intelligence (AI) as the theory and technology of AI is developing, and it's also one of the most popular in them. The goal of ML is not as complex and abstruse as searching for consciousness, candidly speaking, it's to design an kind of algorithm, which can help computer learn. "Learning" does not always mean searching for consciousness, but is more close to find some statistical laws or some patterns. In this view, problem of ML can be divided into two parts: one is design of model, the other is algorithm for search, but the latter is the core of ML.Because of this, this paper consists of two parts:The first part focuses on searching algorithm. Some issues on Evolutionary Computation (EC) are discussed, especially the factors which influence the performance and efficiency of Genetic Algorithm (GA). As Crossover Operation is the key influence factor in them, a new crossover operation is proposed, which named In-and-out Combination crossover operation, simulation results proves the improvement in performance and efficiency via lots of experiments are done for kinds of standard test functions. And then a kind of searching algorithm based on individual-groups is proposed inspired by previous crossover operation, the simulation also presents that they do well than SGA. It can be said that this kind of algorithm is a special sort of Multi-population GA. At last, the concept and significance of "fitness", which is an essential problem in EC, is discussed again in the view of evolution and bionomics. The comparison to some special designing in algorithm is also discussed. Then a measure on adaptive fitness complexity is introduced.In the second part, the design and training of model as another kind of main problems is discussed. Taking the classifier as an example, the reason why multi-learner integration is adopted in the view of performance and efficiency, and methods on integration and training are introduced. Subsequently, in order to use the training sets more efficiently for classifier design on large and complex training sets, a novel strategy of training sets preprocessing which searches for the marginal samples in distance space is proposed. Meanwhile, deficiencies of the methods under some situations are appointed out, and some possible solutions are referred.