GPU-based Parallel Intelligence Algorithms

When it comes to optimization problems,most intelligent optimization algorithms can experience a noticeable deterioration in performance as the scale or the dimensions of problems increase.In reality,for many large-scale overall optimization problems,their scale has continuously increased with technological advances.Hence,how to improve the performance of intelligent algorithms in solving optimization problems of complex functions has attracted much scholarly interest in the field of computation intelligence.Serial intelligent optimization algorithms have difficulties in dealing with large-scale optimization problems.It is therefore necessary to design a rapid,stable and high-precision algorithm.Graphic processing unit(GPU)possesses an obvious advantage over CPU in terms of numerical processing ability and memory bandwidth,whereas it has a low cost and small power consumption.It thus serves as a new alternative for rapidly calculating complex functions.Currently,the computer industry is undergoing a transition from a single application of CPU to a cooperative processing method,which refers to the heterogeneous parallel application of both CPU and GPU.In the latter method,GPU is in charge of numeric processing whilst CPU conducts serial computing such as complex logic and transaction processing.The integrated use of these two can maximize the use of a computer's processing ability.Meanwhile,it not only improves the program performance,but also saves costs and resources,which is a revolutionary improvement.Given that complex functions have various complicated issues,even if efforts have been made to improve the basic algorithms,an optimal solution is still not easy to be obtained.Plus,it can take a long calculation time.Hence,in this paper,we have first modified those basic IAs and designed four improved serial algorithms,which are applied in the GPU devices through the CUDA program.In this way,not only the processing speed is accelerated,but also the accuracy of algorithms gets improved.The paper has mainly focused on the following issues:(1)This paper has brought up a chaotic cuckoo search algorithm based on adaptive mutation strategy(AMSCCS).It utilizes the chaotic mechanism of one-dimension Logistic maps to initialize populations and process newly generated individuals whose value goes above the threshold;through its self-adapting mutation,the algorithm conducts dynamic modification on some of the decomposing variables in accordance with the selection probability.The paper also puts forward a dimension-updating strategy and detection probability of adaptability.Tests conducted on 20 famous complex functions have the following results: When the GPU is NVIDIA Ge Force 310,PAMSCCS has an average speed-up of 3 times and a maximum speed-up of over 8 times compared with the serial algorithm;when the GPU is NVIDIA Ge Force GTX 970,the average speed-up of PAMSCCS is 26 while the maximum speedup is 81.This proves that the performance of algorithm has significantly improved.(2)Regarding the slow speed of convergence and local optimization problem that easily occurred in the progress of using IAs in calculating complex functions with high dimensions,the paper puts forward an improved IWO(IIWO)by properly modifying the basic IWO.To be more specific,the number of each weed's new seeds is set as a fixed parameter,while the initial step size and ultimate step size are adjusted into adaptive step size.Re-initialize the solutions whose value goes beyond the threshold.Meanwhile,by applying this algorithm in GPU,we have come up with an improved PIIWO based on GPU.PIIWO not only improves convergence speed,but also balances the global and local searching ability.The simulation results of calculating over 2010 high-dimension benchmark functions show that,compared with similar algorithms,the IIWO designed in this paper has a better performance,a faster convergence speed and higher accuracy.Compared with the serial algorithm IIWO,the parallel algorithm PIIWO has fewer times of iterative optimization but a better optimization accuracy and an outstanding speed-up.On a low-end GPU,it can get an average three times speed-up while it ca get an average of 22 times and a maximum of over 38 times on a high-end GPU.(3)In order to overcome the deficiencies of genetic algorithms(GAs)which can easily get trapped in local optima and have low accuracy in the later stage,this paper has made certain modification and improvement.Through introducing a two-stage selection operation,random crossover operators and adaptive Gauss mutation operators,we have come up with a parallel quantum-behaved genetic algorithm(PQBGA)based on GPU.The optimization results of 25 complex functions with high dimensions(CEC' 2005)clearly show that the performance of the serial QBGA is superior not only to the basic GAs,but also to other eight famous intelligent optimization algorithms in terms of the convergence accuracy and stability.This proves that the proposed algorithm is both practical and effective.When the QBGA gets applied in the GPU platform,a comparison of QBGA and CPU on two types of GPU shows that the maximum speed-up can reach 18 times and 626 times the original respectively.This means the program's speed and execution efficiency has been remarkably improved,achieving a satisfying speed-up.(4)Spline difference method is used to discrete one-dimensional heat conduction equation into the form of linear equation systems.The system of linear equations is then transformed into an unconstrained optimization problem,before it is ultimately solved by using a parallel hybrid particle swarm optimization(PHPSO)algorithm.The PHPSO is based on CUDA by hybridizing the PSO and conjugate gradient method(CGM).At last,a numerical example is given to illustrate the effectiveness and efficiency of our proposed method.High-end GPU and CPU systems are compared in terms of computation speedup when applied to one-dimensional heat conduction equation.A numerical simulation has been conducted,whose results show that the computational efficiency of the PHPSO algorithm is satisfactory in solving one-dimensional heat conduction equation.The average acceleration ratio of the PHPSO algorithm can reach 12.12 and 13.06 respectively,proving the effectiveness and efficiency of our proposed method.