• View
• Edit
• History
• Print

Parameter Adaptation In The Time-Scale Domain

A method of parameter adaptation has been introduced that can approximate the prediction error of a dynamic model in terms of individual model parameters in the time-scale domain. It achieves this by identifying the pixels (coined as parameter signatures) within the time-scale plane wherein the dynamic effect on the output of individual model parameters is dominant. By attributing the prediction error at these pixels to the error of single parameters, the proposed Parameter Signature Isolation Method (PARSIM) can estimate the error of each parameter separately for iterative parameter estimation. PARSIM has been shown to have an adaptation precision comparable to that of nonlinear least squares (NLS), but to offer the capacity for direct noise compensation in the time-scale domain. PARSIM has also been shown to avoid local minima entrapments that hamper NLS and to be less sensitive to poor data content. The transparency offered by the parameter signatures has also been shown effective in output selection.

The purpose of this presentation is to invite the collaboration of Computer Science colleagues in addressing the numerous theoretical and design issues that remain to be solved. The theoretical issues relate to (i) the convergence characteristics of PARSIM, and the robustness of PARSIM due to poor data content. The design issues correspond to (i) development of potential transformation filters that enhance parameter signature extraction, (ii) the utility of parameter signatures in pruning the nodes of recurrent neural networks, and (iii) the application of PARSIM to adaptation of belief networks.