Global parameter convergence in systems with monotonic parameterization

We consider parameter identification in a class of monotonically parameterized nonlinear systems, one example of which is a neural network. A gradient algorithm is employed to determine the parameter estimates. We determine sufficient conditions on the input under which the estimates converge globally to their true values. We show that new analytical tools that exploit the monotonicity of the underlying nonlinearity and properties of the gradient algorithm can be developed so as to result in global convergence.