Abstract
Liquid State Machines (LSMs) are a computational model of spiking neural networks with recurrent connections in a reservoir. Although they are believed to be biologically more plausible, LSMs have not yet been as successful as other artificial neural networks in solving real world learning problems mainly due to their highly sensitive learning performance to different types of stimuli. To address this issue, a covariance matrix adaptation evolution strategy has been adopted in this paper to optimize the topology and parameters of the LSM, thereby sparing the arduous task of fine tuning the parameters of the LSM for different tasks. The performance of the evolved LSM is demonstrated on three complex real-world pattern classification problems including image recognition and spatio-temporal classification.