A combination of particle swarm optimization and model predictive control on graphics hardware for real-time trajectory planning of the under-actuated nonlinear Acrobot

Abstract

A combination of Model Predictive Control with a Particle Swarm Optimization technique is proposed. The resultant method is able to take advantage of the parallel computation power of graphics hardware to generate swing-up trajectories for the nonlinear & underactuated Acrobot problem in real-time while taking state constraints into account. In order to facilitate this combination, the particle swarm algorithm is improved by making it iterative, this both improves performance and guarantees convergence to a global minimum. The PSO algorithm's convergence rate is further improved by adding a random search term which is dependent on the value of the evaluation function as well as focusing the randomly generated particles around the previous best particle in every iteration of the iPSO algorithm. Simulations are used to investigate this algorithms effectiveness and limitations.