Membrane Computing Based Attribute Weights Allocation Method And Its Application In Pattern Classification

In the analogical learning methods represented by K-nearest neighbor (KNN) andcase-based reasoning (CBR), evaluation of attributes importance is involved, namelythe weight allocation problem. Due to the distribution of attribute weights directlyaffect the calculation of similarity, different weight distribution will get differentsimilarity, and it may also occur big changes on the results of problem solving.Therefore, the research of weight allocation methods has important application valuein field of analogical learning problem solving. In addition, the current weightdistribution methods including subjective methods and objective methods can notensure the rationality of distribution, and it is also difficult to distinguish the relativemerits of these methods among universality. Therefore, it is necessary to explore andresearch the new weight allocation method. Aiming at the problem of optimalallocation of attribute weights in analogy learning methods, this paper proceeds fromthe principle of combination with membrane computing (MC) and weightoptimization mechanism, does a systemic research on the membrane structure andalgorithm of optimizing the weights as well as the contrast experiments inclassification application, and obtains the following results:(1) In order to exert the iterative optimization ability of MC for searching attributeweights, a kind of cell membrane structure is designed which involves theconstruction of surface membrane, secondary surface membrane and several parallelinner membranes, the formulation of membrane rules and region sub-algorithmamong the area. It laid the method for function of weight distribution.(2) With the advantage of parallel computing of MC, the evolution rules aredesigned for parallel inner membranes, such as the selection rule, crossover rule andmutation rule, which ensure the diversity of weight objects for each iteration, theevolution can converge to the optimal solution, and the search scope can be expanded.At the same time, the use of a two-way communication rule will keep the weightobject with largest fitness appeared in the process of evolution and make it participatein the loop iteration, and ensure the weight object output to the secondary surfacemembrane has good performance.(3) In order to avoid the weight falling into the local minimum, a membrane regionalgorithm is obtained by combining the simulated annealing (SA) algorithm with MC.The weight object in secondary subsurface membrane is done annealing operation byperturbation function, and further optimized according to the changes of the fitness.The final optimization result is output to the surface membrane, which promotes theperformance improvement of MC optimization. (4) Combining with the examples of pattern classification based on data, anexperiment platform based on MATLAB-GUI is developed. The contrast ofperformance is studied in both cases, which is defferet parameter settings and thepresence of region sub-algorithm SA in MC. In addition, the method of this paper iscompared with other weight allocation methods by the contrast experiments amongthe classification problems of typical unbalanced data, missing data, medical data andTE process fault diagnosis data, the and results show the effectiveness of the method.In this paper, the weight allocation method based on MC not only improves theperformance of analogical learning methods but also has more advantage than othermethods among application. Experimental results prove that the design of membranestructure and the formulation of membrane rules and region sub-algorithm SA hasrationality, which improves the optimization ability of MC in weight distribution.