Adaptive vs. accommodative neural networks for adaptive system identification: part II

Adaptive neural networks (i.e. NNs with long- and short-term memories), and accommodative neural networks, which are recurrent NNs with fixed weights, are perhaps the most effective paradigms for general and systematic adaptive series-parallel system identification. Adaptive NNs involve less online computation, no poor local minima to fall into, and much more timely and better adaptation than neural networks with all their weights adjusted online. Accommodative NNs do not require online weight adjustment. Part I of this sequence of papers presented in IJCNN\´01 reported that adaptive NNs have much better generalization ability than accommodative NNs in two numerical examples. In this Part, more comparison of the two paradigms is made for series-parallel identification of both deterministic and stochastic plants. Numerical examples show that although adaptive NNs consistently outperform accommodative NNs for generalization, the accommodative NNs have satisfactory generalization performances. However, in an example involving bifurcation and chaos, while the adaptive NN trained on periodic trajectories of the logistic dynamical system tracks accurately its chaotic trajectories, the accommodative NN trained better on the same data fails totally. As reflected in all the examples studied, the variability of the plant outputs seems to directly affect the generalization ability of the accommodative NN. Then an open question is: "How do we measure the variability of the plant outputs to determine whether an accommodative NN has adequate generalization ability for a given application?".