On the performance evaluation of synchronous and asynchronous parallel particle swarm optimisation

In this work, the efficiency (time) and effectivity (fitness) of two parallel variants of Particle Swarm Optimisation (PSO) have been evaluated, the synchronous PSPSO and the asynchronous PAPSO. In this study, an implementation of PAPSO is utilised, which deviates from the master-slave principle. Instead, all particles function as independent workers, competing for the available computing resources. If a particle discovers a new best position, it shares this information with the other particles. Two well-known test functions, the Rosenbrock function and the Rastigin function, were applied for evaluating the efficiency and effectivity of PSPSO and PAPSO. Firstly, versions of the test functions with 10, 30, and 60 dimensions were used. The population size was increased for each dimensionality from 50 to 100 and finally 200 particles. The results of this set of experiments showed that both variants of PSO performed similar regarding to their effectiveness of finding the optimum solutions. The computing time used by PAPSO, on the other hand, is significantly smaller than the computing time needed by PSPSO. On average the PAPSO was 69.1 % faster than the PSPSO on the Rosenbrock function and 90.3 % faster on the Rastigin function. In a second set of simulations, the maximum waiting time was varied from 5 ms to 1,000 ms. It is shown for both algorithms, that the average computing time rises linearly with the maximum waiting time.