Implementation of a multi-layer perceptron for a non-linear control problem
A 1-hidden-layer multilayer perception (MLP) is used to provide a non-linear continuous multi-variable mapping with 42 inputs and 13 outputs. The problem is that of extracting information from experimental signals with a bandwidth of many kilohertz, given an exact model of the inverse mapping of the problem, but no explicit form of the required forward mapping. The MLP has been trained to provide this mapping by learning on 500 input-output examples. The success of the off-line solution to this problem using an MLP has led to an examination of the robustness of the MLP to different noise sources. The MLP is found to be more robust than an approximate linear mapping of the same problem. 12 bits of resolution in the weights are necessary to avoid a significant loss of precision. The practical implementation of large analogue weight matrices using DAC-mulipliers and a simple segmented sigmoid is also presented. A General Adaptive Recipe (GAR) for improving the performance of conventional back-propagation uses an adaptive step length and both the bias terms and the initial weight seeding are determined by the network size. The GAR is found to be robust and much faster than conventional back-propagation.
Bibliographic Reference: Report: LRP 398/90 EN (1990) 46 pp.
Availability: Available from Confédération Suisse, Centre de Recherches en Physique des Plasmas, Ecole Polytechnique Fédérale de Lausanne, 21 avenue des Bains, 1007 Lausanne (CH)
Record Number: 199110146 / Last updated on: 1994-12-02
Original language: en
Available languages: en