Uniform Initialization in Response Space for PSO and its Applications

Particle swarm optimization (PSO) is widely used in the parameter estimation for complex models, which is the key to establishing a mathematical model. However, convergence to the local optimal easily occurs in PSO. A substantial amount of researches have been conducted to improve the evolutionary process of PSO; nonetheless, the study on initialization method is relatively limited. Generally, initial particle positions are uniformly distributed in the parameter space but unevenly distributed in the response space for nonlinear models, which can hinder optimization. In this paper, a novel initialization method for PSO with an uninformative prior of parameters in the model (UPPSO) is proposed. The method initializes particle positions uniformly distributed in the response space and makes the effect of particle velocity uniform in the response space. The UPPSO and other five algorithms were applied to estimate parameters for three different polymer models, that is, viscosity and PVT (pressure-volume-temperature) models which are very important and must be estimated for each type of polymers. In comparison with other algorithms, the optimization capacity of UPPSO for each model was ranked second, and UPPSO was particularly outstanding in obtaining the optimal parameter. Moreover, UPPSO was also competitive in computation performance. In general, UPPSO is conducive to optimization, efficient and adaptable, and it can also be generalized to other swarm intelligence algorithms, such as the differential evolution (DE) and artificial bee colony (ABC), etc.