Inctemental Learning Algorithm Of Fuzzy Semantic Cell

Concept is the knowledge unit formed by designated features.Concept representation,concept learning and concept computation are crucial to artificial intelligence focused on simulating human mind.Because of its fuzziness,concept can't be simply quantified.The fuzzy semantic cell model solves this problem,L=<P,d,8>contains a prototype to describe typical samples and a distribution to describe distance between typical samples and other samples.This is focused on concept learning using the fuzzy semantic cell model.Current fuzzy semantic learning algorithms all work in off-line mode,so the learning result heavily depends on the original set of samples.Once the algorithm comes to an end,cells will not be able to be updated by a new sample,we must add the new sample to the set and restart the whole algorithm.Off-line mode limits the learning ability of the fuzzy semantic cell model,making it impossible to handle abundant and changeful data in real applications.Therefore,we propose an incremental learning algorithm about fuzzy semantic cells to solve the drawbacks of the existing algorithms.This thesis's work is divided into two parts.In the first part,we improve the existing algorithm of fuzzy semantic cell learning to make it better to provide initial fuzzy semantic cells for incremental learning.In consideration of the feasibility of the concept representation,the computational complexity and the potential replication of the algorithm on exponential family members,we chose the exponential distribution as the distribution of cells.And for the existing multiple fuzzy semantic cells learning algorithm,we adjust its objective function so that it aims at "extracting accurate concepts" rather than "accurately describing the sample set".Since initial cells selection and regulatory factor have a large effect on the result of the algorithm,we make a logical analysis and optimize it.In the second part,we put forward an incremental learning algorithm about fuzzy semantic cells.We propose seven rules for fuzzy semantic cells updating,merging and creating to handle learning conditions from single fuzzy sematic cell to multiple fuzzy semantic cells,one sample to multiple samples,non-supervision to supervision.We also propose two ideas about how to update cells.One we named approximation idea,which follows principles of cell similarity,sample membership enhancement and semantic meaning definition,is a conservative way to update a cell.The other is generation idea,which obeys principles of justifiable granularity and maximum fuzzy entropy just as fuzzy semantic cell learning algorithm does,is an update way paying more attention to the whole samples.Experimental results proves that our incremental learning algorithm of fuzzy semantic cell is robust.It can well cope with all kinds of situations caused by new sample adding,and the fuzzy semantic cells updated by this incremental learning algorithm can still maintain its excellent performance on concept representation.