Reinforcement learning using swarm intelligence-trained neural networks

Abstract

This article proposes a new reinforcement learning method using the swarm intelligence-trained neural network (SWINN) to generate solutions to various real-world problems efficiently. The swarm intelligence algorithm, particle swarm optimisation (PSO), is combined with a training resource allocator (TRA) in SWINN for specific problems. TRA, as a heuristic global search method, controls the allocation of training resources to different candidate topologies of artificial neural networks (ANNs) to expedite the system convergence, while PSO is applied as a local search algorithm to adjust the ANNs connection weights. To evaluate the performance of the SWINN algorithm, two reinforcement learning case studies: the double pole balance (a.k.a. double inverted pendulum) problem and a mobile robot localisation problem are conducted. Extensive simulation results successfully demonstrate that SWINN offers performance that is competitive with modern neuroevolutionary techniques, and is viable for real-world problems.