Parallel Continuous Ant Colony Optimization Algorithm Based On GPU And Its Application Research

Ant Colony Optimization (ACO) is a swarm intelligence based optimization algorithm which is inspired by the foraging behavior of real ant colony. ACO is initially proposed for the optimization problems in discrete domains. But in actual project, there exist a lot of continuous optimization problems that is unsuitable to the basic ACO. On the other hand, in a practical application, when the problem has high accuracy requiring and high dimension, ACO has high time consumption. Nowadays, Graphics Processing Unit (GPU) has a much stronger floating computing ability than CPU and its processor is designed to be parallel architecture. GPU provides a powerful hardware platform for running real parallel program. This paper studies the method of expanding ACO into continuous domains and makes a parallel design based on GPU for continuous ACO to improve the execution efficiency. The researches in this paper include:(1) A continuous ACO is introduced for solving continuous optimization problem, following the basic framework of ACO. Its principle and mechanism in detail are discussed. The algorithm is tested on benchmark functions and compared with several classic algorithms. The simulation results prove its stronger optimization ability.(2) A parallel implementation of continuous ACO on GPU is proposed. To fully use the massively parallel computation capabilities offered by Compute Unified Device Architecture(CUDA), the algorithm is divided into three kernel functions. The implementation details and core codes of kernel functions are present. The performance analysis between parallel version on GPU and serial version on CPU are given. The parallel continuous ACO achieve satisfactory speed-up ratio.(3) The Radio Frequency Identification Device (RFID) signal strength distribution modeling problem is studied. Feedforward neural network is used to build a signal distribution model and the parallel continuous ACO is employed to optimize the parameters of neural network. The experiment system is constructed with RFID device to collect the signal strength data for training network. According to the experiment results, the model can mainly reflects the fact of signal distribution. The parallel continuous ACO reduces much execution time.The main innovation points of paper include:(1) A new ACO in continuous domain is brought to make ACO suitable for continuous optimization problem. A parallel implementation of the algorithm on GPU is made and the execution efficiency is much improved.(2) The continuous ACO is applied to optimize the parameters of neural network. The model of RFID signal strength distribution is built and it has a certain application value.