“…To overcome this problem, an ensemble technique is utilized. 31 In this technique, Q sets of different nonlinear models F 1,2 ͑i͒ ͑i =1, ... ,Q͒ are collected, each of which can be obtained by the same procedure as ͑B2͒ except that a different initial condition, which is generated randomly, is set for the parameter estimation. This results in Q sets of neural networks with the same architecture with different parameter values W. An ensemble average is then taken as F 1,2 = ͑1 / Q͚͒ i=1 Q F 1,2 ͑i͒ .…”