From Noise-Corrupted Data
T. Troudet, Walter C. Merrill · NASA Technical Reports Server (NASA) · 1990
Summary The ability of feed-forward neural network architectures tolearn continuous-valued mappings in the presence of noise wasdemonstrated in relation to parameter identification and real-time adaptive control applications. An error function wasintroduced to help optimize parameter values such as numberof training iterations, observation time, sampling rate, andscaling of the control signal. The learning performancedepended essentially on the degree of embodiment of thecontrol law in the training data set and on the degree ofuniformity of the probability distribution function of the datathat are presented to the net during a training sequence. Whena control law was corrupted by noise, the fluctuations of thetraining data biased the probability distribution function of thetraining data sequence. Only if the noise contamination isminimized and the degree of embodiment of the control lawis maximized, can the neural net develop a good internalrepresentation of the mapping and be used as a neurocontroller.A multilayer net was trained with back-error-propagation tocontrol a cart-pole system for linear and nonlinear control lawsin the presence of data processing noise and measurementnoise. The neurocontroller exhibited noise-filtering propertiesand was found to operate more smoothly than the teacher inthe presence of measurement noise.