The back propagation method is well known as a supervised learning rule of a neural network. In this paper, a new learning rule is proposed where the output error vector is adjusted to zero by correcting two kinds of vectors, the one is weighting vectors (matrix) and the other is an input vector of the layer. The corrected input vector has a role of a tentative teacher of the following layer. In this way, the output error is propagated backward, and is partly corrected by each weighting vector. Computatinal method is also presented for a matrix inversion which is required in the proposed method and nonsingularity of the matix is discussed. Simulation result of the Exclusive-OR problem shows the effectiveness of the proposed method.
Progress in recent sensor technology has inspired us to start research on recovering an object surface from depth data together with additional sensory information, i.e. data on its differential geometry or a priori knowledge about its smoothness. Any method for the recovery with both classes of the additional information, however, has not been proposed yet. Such a problem can be generalized to that of function approximation using its different order derivatives, which covers many important applications. To solve the problems generally, we propose a method of fusing all the information as follows : (1) To evaluate the smoothness and penalties for the derivative data of each order, we define a functional. (2) All the information are related to one another by restricting the fused results in the set of stationary functions of the functional. By the fusion, a new approximation is realized as the stationary function that minimizes all penalties simultaneously.
This paper gives a controller design method of 2 mass-spring system which satisfies the given H∞ control performance and pole assignment in the prescribed region in the presence of physical parameter perturbations. In addition, its effectiveness is demonstrated by experiments.
It is shown that combining multivariate analysis with neural networks is useful for solving problems with complex interaction such as a vehicle aerodynamics of lift. This method leads us to various valuable observations. In this paper, we show the effectiveness of this method. Regarding some of the past insufficiencies of solving problems by multivariate analysis method or neural networks only, there are conditions such as (1) difficult to collect data, (2) mere knowhow, (3) predicting complex interaction. But by combining neural networks method with multivariate analysis, we can solve the problems efficiently. In addition, solving this problems by combining neural networks method with multivariate analysis, new detections about the works of hidden layers in neural networks were found.
For a plant consisting of a lumped non-minimum phase part and an output delay, we discuss LTR (Loop Transfer Recovery) techniques for designing a predictor-based LQG controller. We focus our attention to the feedback property achieved at the plant output side. We show that a feedback property recovered by the formal application of the conventional LTR procedure can be achieved by the partial LTR technique which has a clear system-theoretic meaning. This fact provides a theoretical justification of the formal application of the conventional LTR technique to a plant including finite unstable zeros and a time delay. In addition, we point out that the partial LTR technique provides more freedom in shaping target feedback properties. We propose a simple technique exploiting this freedom. A numerical example is given to illustrate usefulness of the proposed technique.