Reservoir Computing optimization with a hybrid method

Reservoir Computing (RC) is a paradigm of artificial neural networks with important applications in the real world. RC uses similar architecture to recurrent networks without the difficulty of training the network hidden layer (reservoir). However, RC can be computationally expensive and various parameters influence its efficiency, making it necessary to search for alternatives to increase its capacity. This work aims to use a hybrid algorithm between a PSO (Particle Swarm Optimization) extension and Simulated Annealing for optimize the global parameters, architecture and weights of RC, in time series forecasting. The results showed that the Reservoir Computing optimization with the hybrid algorithm achieved satisfactory performance in all databases investigated and outperformed original APSO (Adaptive Particle Swarm Optimization) in some of them.