Artificial Neural Network Structure Learning Algorithm And Problem Solving

Algorithm's unable to compute the structural parameters of neural networks places restrictions on many existing learning algorithms in many ways such as the slow learning velocity, the weak error-tolerance and so on. In addition, it is also restricted using already given neural network model to perform particular problem-solving computation. For example, the solution is often inaccurate. All the above indicates that we should determine the neural network's structure by the specific problem.In this paper, the basic conception and typical learning algorithms with their properties are first described briefly, then the main creative results are proposed as follows:(1) A learning algorithm of forward neural networks for general binary mapping problems is presented. This algorithm gives a new kind of method to divide the training points into several areas by computation of hyper-planes. It constructs the hidden layer neural networks correspondingly and makes it possible to carry out the mapping described by the training samples only by constructing one output layer to it This algorithm converges faster than BP and SC algorithm, adapts strongly to the degree of the sample data's distribution and concentration, has strong capability of error-tolerance and is facilitated to modem VLSI implementation.(2) Two feedback associative memory network learning algorithms called competing-classifying and adaptive-competing-classifying algorithm and one feed forward associative memory network learning algorithm called forward-comparing algorithm are presented. The three algorithms are named for short CC, ACC and FC algorithm respectively. The neural network conducted by CC-algorithm can store p sample patterns via p+n neurons, so the capacity of the network approach the best. For an arbitrary input pattern, CC network will steadily output the sample pattern which has the least Hamming-distance from it, so the algorithm's error-tolerance capability is optimal. ACC algorithm makes the linking-weights change adaptively during the competition procedure, and makes the corresponding network converge much faster than CC network. ACC network has the same error-tolerance capability as CC network and its capacity nears half of that of CC network. FC algorithm constructs feed forward associative memory neural networks. Each computation can complete in only one step. FC network's error-tolerance capability is identical to CC network but its capacity turns to 0(N)^1/2 where N is the number of neutrons.(3) In the study of real-time neural networks computation we bring forward a kind of neural networks construction method for solving the shortest path problems and prove that the neural network constructed according to the method has only one equivalence-point. The network can converge to this equivalence-point from arbitrary initial point and this equivalence-point corresponds exactly all the shortest path pairs of the graph. This is a new try of solving non-NP hard problems via neural networks...