Classification Of Complex Structure Patterns Based On Recursive Neural Networks

With the development of pattern recognition, artificial intelligence and machine learning, traditional statistical classification and recognition methods can't classify and recognize some complex problems effectively. It is reported that, more and more research objects have been transformed from high-dimension vectors in the non-structured domains to trees or DOAGs structured vectors in the structured domains. The main research object is structured data, and classifying and recognizing the whole pattern by classifying and recognizing its sub patterns.This article does lots of research on supervised recursive neural networks and unsupervised recursive neural networks which can deal with structured data. The research of recursive neural networks is based on the research of normal neural networks. This article first describes normal neural networks and then generalizes them to the structured domains just as the order of research. When dealing with forward neural networks, this article carries experiments on BP, LMBP and PSO training algorithms. When dealing with supervised recursive neural networks, this article describes the BPTS training algorithm as well as its improved algorithms and the best PSO algorithm in training forward neural network is also used to train recursive neural networks. When dealing with SOM, this article discusses growing SOM and parameterless SOM respectively, then proposes parameterless growing SOM based on these. When dealing with unsupervised recursive neural networks, this article generalizes the basic SOM, growing SOM, parameterless SOM and parameterless growing SOM to the structured domains.The main contributions are listed in the following contents:1. This article proposes new PSO algorithms based on "Global Best" factor which can speed up the convergence greatly when training neural networks compared with BP and other particle swarm algorithms. Furthermore, with the experiences of PSO algorithms in the application of function optimization, this article concludes a more generalized particle swarm algorithm by merging the proposed algorithms and the algorithms existed together. The generalized particle swarm algorithm gives a suggestion of how to apply particle swarm algorithm into other domains. This article proposes a quick training algorithm on supervised recursive neural networks based on PSO algorithms by applying the PSO algorithm which performs best in training neural...