Research On Memristor Network-Based Typical Swarm Intelligence Algorithm

Swarm intelligence is inspired by the intelligent behavior of biological swarms.Ant colony optimization algorithm and genetic algorithm are two typical swarm intelligence algorithms.Benefiting from its distribution,swarm intelligence can realize parallel computing by hardware circuits such as memristor networks.Memristor is a non-volatile,programmable circuit device,which has many excellent physical characteristics.Compared with the traditional computing model with separation of storage and calculation,the memristor can realize the in-memory computing model(storage and calculation are completed in the same place),eliminating the transmission cost between the memory and the calculation unit.Therefore,memristor network is applied to parallel computing and in-memory computing deployment of algorithms such as deep neural networks and swarm intelligence.In this context,this thesis focuses on the ant colony optimization algorithm and genetic algorithm based on memristor network and their applications in data dimensionality reduction methods.The main work and contribution are summarized as follows(1)Ant colony optimization algorithm based on memristor networkThe dynamic change process of conductivity value in memristor is similar to pheromone in ACO.In terms of models,a voltage controlled memristor with relaxation factor is proposed,and its relationship with ant colony optimization algorithm is established.In terms of experiments,the solution based on memristor circuit and the solution based on ant colony optimization algorithm are constructed respectively,and it is found that the memristor value and pheromone have similar change curve.In terms of circuit,a memristor network deployment of ant colony optimization algorithm based on memristor cross array is designed to reduce the time complexity of the algorithm from square level to linear level.(2)Genetic algorithm based on memristor networkMatrix computing has high hardware parallelism,and memristor network can realize the parallel and in-memory computing matrix computing model.Considering that the population of genetic algorithm can be represented by matrix and the operators can be calculated by vector and matrix,a matrix friendly genetic algorithm is proposed.Compared with the experimental results of the baseline genetic algorithm,the matrix friendly genetic algorithm has better convergence results,faster convergence speed,and can save nearly 2/3 of the running time.The memristor network deployment of matrix friendly genetic algorithm is designed,and reduce the time complexity of the operator to linear level which related to the number of genes.(3)Feature extraction based on memristor network swarm intelligence algorithmThe heuristic search capability of the swarm intelligence algorithm can effectively deal with the combination optimization problem faced by feature extraction,and the memristor network can improve the computational efficiency.The image edge feature extraction is realized by the ant colony optimization algorithm based on the memristor network.Compared with the traditional operator method and the deep neural network method,the image noise and the amount of detail can be better balanced.On the Sonar data set,the feature selection implemented by the genetic algorithm based on the memristor network uses about 24% of the number of features,which can improve the accuracy of the logistic regression model by 10.1%.