Takayuki Yamada

Learning control using neural networks

The basic features of the learning-type neural network (NN) controller are clarified. Analytical and experimental results show its stability, convergence and generalization ability compared with the adaptive-type NN and conventional learning control. As an application of the learning-type NN, a nonlinear optimum regulator is presented whose learning ability can obtain optimum conditions without solving a difficult Riccati equation. Moreover, it can be applied to a nonlinear control system because of its nonlinear mapping ability, although the conventional optimum regulator can only be applied to a linear system. Finally task planning is proposed in terms of skill acquisition using the learning-type NN, which implies the possibility of making an interface with an upper symbolic-level control.<<ETX>>

Possibility of neural networks controller for robot manipulators

NN (neural network) controller characteristics are clarified by comparison with the adaptive control theory. The authors explain the classification of the NN controller architecture and the dynamic NN structure. A comparison between the NN controller and the adaptive controller shows that the framework of a linear two-layer NN controller is the same as that of the adaptive controller, and that the nonlinear three-layer NN (PDP, or parallel distributed processing type) is a nonlinear extension of the adaptive control. The stability characteristics of the NN control system, which shows the robustness effect of the generalized delta rule, the plant and the NN mapping function, are treated. Finally, NN controller experiments are demonstrated using a force control servomechanism. Experimental results suggest that the nonlinear sigmoid function of the NN can compensate for the nonlinear plant effect.<<ETX>>

Force control of servomechanism using adaptive control

It is shown that a servomechanism with only position/velocity functions can obtain a force control function using the adaptive theory (model-referenced adaptive system and model-referenced control). Experimental results of both the servomechanism and the single-degree-of-freedom force control of the manipulator show that this method leads to suitable force control when the object stiffness changes greatly. The advantage of this method is that a force control function is achieved by implementing software to identify the object stiffness without changing hardware. >

Nonlinear neural network controller for dynamic system

A learning type of controller using a neural network is proposed and compared with a conventional controller. The learning neural network controller can use not only quadratic error but also a more general cost function. A practical design method is proposed, and the advantages of the learning type of neural network controller in comparison with the adaptive type are discussed. Simulated and experimental results confirm the realization of nonlinear optimal control using the proposed controller.<<ETX>>

An extension of neural network direct controller

Many studies have been undertaken in order to apply both the flexibility and learning ability of neural networks to dynamic system controllers. The paper proposes several kinds of neural network direct controllers as extensions to that previously proposed. They are an open loop type controller, in which the plant output information is not needed to realize the dynamics and the neural network direct controller has a feedback loop (a feedforward/feedback type). Theoretical studies guarantee the stability of the proposed controllers. Simulation and experimental results confirm both the realization and the characteristics of the controllers.<<ETX>>

Remarks on an adaptive type self-tuning controller using neural networks

The authors propose an adaptive type of self-tuning controller which offers the possibility of enhanced robustness. This controller does not use a desired feedback gain predetermined by conventional control theories as a teaching signal. Simulated results using a second-order plant show the nonlinear neural network effect for the nonlinear plant. They also show that a controller using a linear neural network is better than one using a nonlinear neural network in the region of small nonlinear and parasite terms. Experimental results using a force confirmed that the controller is useful for an actual system.<<ETX>>

On the characteristics of the robot manipulator controller using neural networks

Several robot controller schemes using neural networks are proposed. To extract advantages from the neural network controller, it is necessary to clarify the capabilities of the neural networks to express a nonlinear function as well as the robustness of the controller. For this purpose, an inverse kinematics problem and a force control problem are presented. The characteristics of the neural network are compared to adaptive control for the robot controller using the force control problem as an example.<<ETX>>

Remarks on hybrid neural network controller using different convergence speeds

A neural network requires the partial derivative of a plant output with regard to its input. However, it is unknown for an unknown nonlinear plant. This paper proposes a hybrid neural network controller which overcomes this problem and which compensates online neural networks for plant fluctuation by using an identifier and a controller with different convergence speeds.

Digital control stability improvement of master-slave manipulator system

Proposes a digital control method for master-slave manipulator systems which can realize high-fidelity telemanipulation. In the discrete-time system, the time-delay of sensor signals and the zero order hold effect of command signals on actuators should be considered to make the systems stable. The authors first analyze the dynamic control method of a master-slave system in discrete-time for the stability problem, which can realize high-fidelity telemanipulation in continuous-time. Secondly, using the results of the stability analysis, the robust control scheme for the master-slave system is proposed, and the validity of this scheme is finally confirmed by simulation.<<ETX>>

Auto-tuning of feedback gains using a neural network for a small tunnelling robot

Describes the auto-tuning of feedback gains for a small tunnelling robot. The authors (1989) have already proposed the directional control method that the head angle of the control input is the sum of the deviation multiplied by feedback gain Kp and the angular deviation multiplied by feedback gain Ka. In this paper, they use a neural network to obtain feedback gains Kp and Ka. The inputs of the neural network are an initial deviation and an initial angular deviation. The outputs of the neural network are the feedback gains Kp and Ka. This neural network learns from the deviation errors. The optimum gains obtained by the proposed method agreed with the optimum gain obtained by trial and error. The neural network which can apply to any initial deviations were formed by using plural initial deviations in learning. Moreover, this method can tune the optimum gains to any design line. The results showed the validity of the proposed auto-tuning method.<<ETX>>