Showing papers by "Keigo Watanabe published in 1991"

Learning algorithms of layered neural networks via extended kalman filters

Abstract: Learning algorithms are described for layered feedforward type neural networks, in which a unit generates a real-valued output through a logistic function. The problem of adjusting the weights of internal hidden units can be regarded as a problem of estimating (or identifying) constant parametes with a non-linear observation equation. The present algorithm based on (he extended Kalman filter has just the time-varying learning rate, while the well-known back-propagation (or generalized delta rule) algorithm based on gradient descent has a constant learning rate. From some simulation examples it is shown that when a sufficiently trained network is desired, the learning speed of the proposed algorithm is faster than that of the traditional back-propagation algorithm.

Adaptive generalized predictive control using a state-space approach

Abstract: A state-space approach is described for designing adaptive generalized predictive controls for single-input/single-output CARIMA systems having unknown plant parameters. An observer canonical innovation model is used for designing the j-step-ahead output predictors, in which the direct output method, that is based on using the plant output directly, and the output deviation method, that is based on using the deviatory output, are considered. The unknown parameters included in such long-range predictors are identified by employing an extended least-squares method with an exponential weighting. Two particular control systems are developed, based on these principles, and their performances are compared. >

An interative learning control scheme using the weighted least-squares method

Abstract: An iterative learning control scheme is described for linear discrete-time systems. A weighted least-squares criterion of learning error is optimized to obtain a unique control gain for a case when the number of sampling is relatively small. It is then shown that algorithmic convergence can be readily guaranteed, because the present learning rule consists of a steady-state Kalman filter. By paying attention to the sparse system structure for the system's impulse response model, we further derive a suboptimal iterative learning control for a practical case when the number of sampling is large.

Impact sound control by learning control

Abstract: Impact sound is emitted when a collision occurs between objects. The authors propose a robotic control method to reproduce the impact sound emitted by the collision between the endpoint of the robotic manipulator and the object. The feedback control of impact sound is very difficult, because the impact phenomenon occurs in a very short period of time. Instead of a feedback control method, a learning control algorithm is proposed to control the impact phenomena. The learning control algorithm modifies the reference signal to the servo controller of the robot manipulator so that the desired impact sound is reproduced. The algorithm is based on the optimization of the least-squares criterion of learning error and does not require the dynamic model of the impact phenomena. The algorithm is applied to the planar manipulator with one degree of freedom driven by a pneumatic actuator. Experimental results illustrate the effectiveness of the proposed algorithm. >