Abstract
This paper examines Echo State Network, a reservoir computer, performanceusing four different benchmark problems, then proposes heuristics or rules ofthumb for configuring the architecture, as well as the selection of parametersand their values, which are applicable to problems within the same domain, tohelp serve to fill the experience gap needed by those entering this field ofstudy. The influence of various parameter selections and their valueadjustments, as well as architectural changes made to an Echo State Network, apowerful recurrent neural network configured as a reservoir computer, can bechallenging to fully comprehend without experience in the field, and even somehyperparameter optimization algorithms may have difficulty adjusting parametervalues without proper manual selections made first. Therefore, it is imperativeto understand the effects of parameters and their value selection on Echo StateNetwork architecture performance for a successful build. Thus, to address therequirement for an extensive background in Echo State Network architecture, aswell as examine how Echo State Network performance is affected with respect tovariations in architecture, design, and parameter selection and values, aseries of benchmark tasks representing different problem domains, includingtime series prediction, pattern generation, chaotic system prediction, and timeseries classification, were modeled and experimented on to show the impact onthe performance of Echo State Network.