Research And Application Of Artificial Neural Network Based On Particle Swarm Optimization Algorithm Using GPU

Artificial neural network(ANN) is a massively parallel processor and can be used to solve complex optimization problems by highly interconnected neurons. ANN model has been applied to electromagnetic and measurement fields due to its advantages of excellent abilities of learning and generalization. Particle swarm optimization(PSO) algorithm has gradually replaced the common error back propagation(BP) algorithm and been applied to ANN training(PSO-ANN) due to its easy implementation and strong abilities of convergence and global search. However, PSO-ANN needs long computing time especially for large scale problems. Parallel optimization is an effective way to solve this problem. Besides ANN’s data parallelization and node parallelization, PSO’s natural particle behavior parallelization is in PSO-ANN. There are many parallel ways to accelerate PSO algorithm. Compared with computer cluster, multi-core CPU or other professional parallel devices like FPGA, graphic processing unit(GPU) has the most significant advantages in hardware cost. Since the NVIDIA company introduced the compute unified device architecture(CUDA) in 2007, CUDA C has become the most popular GPU programming language due to its excellent programmability.Ground on the existing research of GPU-based parallel PSO algorithm and GPU-based BP-ANN algorithm using data parallelization, we design the CUDA-based parallel PSO-ANN scheme and parallel PSO-BP-ANN scheme, and use these schemes to fast model the resonant frequency of microstrip antennas(MSA) and the direction of arrival(DOA) estimation in this paper. Experimental results show that the modeling error is superior to the results of the existing researches within a short time.The main researches of this paper are as follows:(1) To deal with the unreasonable aspect of traditional performance index, we propose the concept of “Effective Speedup” to measure the achievement of GPU-based parallel PSO algorithm and CPU-based sequential PSO algorithm. Benchmark test functions are tested and analyzed.(2) Ground on the existing research of GPU-based parallel PSO algorithm, we design the CUDA-based parallel PSO-ANN scheme to fast model a test function approximation.(3) We use the designed GPU-based parallel PSO-ANN scheme to fast model the resonant frequency of both rectangular MSA and circular MSA, and analyze the modeling speed and modeling error.(4) Ground on the designed GPU-based parallel PSO-ANN algorithm and the existing research of GPU-based BP-ANN algorithm using data parallelization, we design the CUDA-based parallel PSO-BP-ANN scheme, use the designed scheme to fast model the DOA estimation, and analyze the modeling speed and modeling error.