| In recent years,due to the rapid development of computing power,Deep Learning(DL)algo-rithm has been applied for a variety of scenarios.Among them,the Convolutional Neural Network(CNN)is proposed to solve many complicated problems in the image processing tasks,such as image classification and image segmentation.However,for a classification task,the performance depends not only on the neural network architecture but also on dataset.With the rapid increase of data,it is necessary to develop the automatic search algorithm of neural network.Aiming at the classification tasks,this thesis mainly studied the evolutionary algorithm of convolutional Neural network based on NeuroEvolution augmentation Topology(NEAT),and implemented the general search strategy algorithm based on a Neural Architecture Search(NAS)data set,NASbench-101.Firstly,this thesis introduces the basics of neural networks in detail,including the basics of con-volutional neural networks(CNN)and recurrent neural networks(RNN),whose structures largely determine the search space of NAS algorithms.Then,the value based Q-learning and DQN algo-rithms of Reinforcement Learning(RL)as well as the principle of genetic algorithm are introduced in this thesis,and they are more commonly employed as search strategies in the current NAS re-search.The above contents are necessary theoretical basis for subsequent research.This thesis proposes C-NEAT algorithm based on NEAT algorithm for classification tasks.Firstly,we mainly introduces the basics of neural evolution algorithm and NEAT algorithm,and we elaborates on the genetic encoding scheme of NEAT algorithm for neural network,as well as the principle of crossover mutation operator.The C-NEAT algorithm extends the minimum encoding scale of an individual network from the neuron to a specific layer.Further more,this thesis expands the encoding expression of NEAT node genes and defines the specific mapping relationship between NEAT graph and CNN.Finally,the C-NEAT algorithm is applied for intrusion detection data set,KDD99,and the standard image data set,CIFAR10.The simulation results varifies the effectiveness of the C-NEAT algorithm.In order to evaluate the effectiveness of various NAS algorithms,this thesis studied the NASBench-101 dataset,which is a novel NAS dataset,and we apply Monte Carlo Tree Search(MCTS)Search to solve the NAS task in the NASBench-101 dataset.We introduces the network architecture de-sign,cell structure encoding scheme and the performance evaluator of NASbench-101 in detail.The presence of NASbench-101 greatly reduced the limited computing power problem for NAS researchers.Then,this thesis introduces several general NAS algorithms,including random search algorithm,generalized evolutionary algorithm and reinforcement learning search.Later,We im-plements the above algorithm on NASbench-101 dataset.Then,this thesis redefines the state space and action space for NAS in reinforcement learning,and uses Monte Carlo Tree Search(MCTS)algorithm to solve the predefined search problem,which is applied to the Nasbench-101 data set.Final simulation results show the effectiveness of the MCTS algorithm which is proposed in this thesis. |
