Separable Least Squares Identification of a Parallel Cascade Model of Human Ankle Stiffness

The identification of a dynamic, nonlinear model of human ankle stiffness is considered in a minimum mean squared error framework. The model consists of two parallel pathways, one representing the intrinsic dynamics, the other representing the reflex contribution to the stiffness. The model is shown to be linear in all of its parameters, except for those used to describe a single static nonlinearity in the reflex pathway. A separable least squares optimization algorithm is developed which takes advantage of this structure. This new algorithm is applied to experimental stretch reflex data, and the results compared to the current state-of-the-art algorithm, an iterative technique which fits the two pathways alternately. The relative merits of the two approaches are discussed. Keywords— nonlinear system identification, mean squared optimization, separable least squares, stretch reflex dynamics.