In order to capture uncertainty of the control system based on input-output data, multiple data are necessary. However, model sets determined based on multiple input-output data tend to be conservative. Moreover, the conservatism depends on the degree of freedom in parameterizations of model sets to be determined. In this paper, we show the existence condition and the parameterization of model sets which have much degree of freedom and are not inconsistent with given input-output data. Moreover, we propose a method to obtain the smallest model set by minimizing volume of model sets. The minimization problem can be reduced to linear matrix inequalities (LMIs) and can be solved efficiently.

抄録全体を表示

In this paper, we propose a design method of repetitive controllers. Our method is based on the factorization approach that consists of the coprime factorization of (augmented) plants and the solutions to so-called Diophantine equation. We do not use the small gain theorem. In our previous paper, we propose a design method for minimum-phase plants. Therefore, in this paper, we extend our method to nonminimum-phase plants.

抄録全体を表示

In this paper, we discuss a simple and robust controller for flexible large space structures based on a distributed parameter model. As a typical example of large space structures, we consider N flexible beams connected by springs. The flexible beams and the springs can be regarded as an element of the structure with the distributed flexibility and a connective part with lumped flexibility, respectively. We derive dynamic equations of the distributed parameter system by means of the Hamilton's principle. We introduce a Lyapunov function related to the total energy of the distributed parameter system and derive a simple sensor output feedback control law (PDSS feedback control law). Using the separation of variables and the invariance principle, we prove the asymptotic stability of the closed-loop system and the convergence property to the desired stationary state. As we don't need an approximated finite-dimensional model at the controller design phase, the derived controller based on the original distributed parameter system is robust and simple. In order to demonstrate the validity of the proposed controller, experiments have been carried out.

抄録全体を表示

This paper describes a new mechanical element that varies its stiffness to internal air pressure. The element, named Particle Mechanical Constraint (PMC), is a soft vinyl tube embedded with Styrofoam beads. Its shape varies freely with compression, elongation, bending and twisting. By exhausting the inside air, the PMC solidifies in an arbitrary shape and constrains all degrees-of-freedom. Stiffness and viscosity are adjustable via the inside air pressure. The PMC is light and safe, and therefore can be used as a wearable device. The application of a PMC to a wearable haptic display is described in this paper.

抄録全体を表示

This paper presents a joint synthesis strategy of a controller and a fault detector, and its application to a motor drive control system. The influence of a fault in power electronic devices is analyzed and simulated. Some experimental results demonstrate the effectiveness of the proposed synthesis method.

抄録全体を表示

In real world applications, learning algorithms often have to handle several dozens of actions, which have some distance metrics. Epsilon-greedy or Boltzmann distribution exploration strategies, which have been applied for Q-learning or SARSA, are very popular, simple and effective in the problems that have a few actions, however, the efficiency would decrease when the number of actions is increased. We propose a policy function representation that consists of a stochastic binary decision tree, and we apply it to an actor-critic algorithm for the problems that have enormous similar actions. Simulation results show the increase of the actions does not affect learning curves of the proposed method at all.

抄録全体を表示

The storage function method (SFM) has been the main method of rainfallrunoff forecasting used in Korea. However, it has a major drawback in that the SFM's parameters are difficult to calibrate exactly. They have large degrees of uncertainty and may also be time varying. To cope with these difficulties, we present an adaptive storage function method (ASFM). Under the assumption that a small basin can be modelled by a single watershed and channel pair, a multiple model adaptive estimation (MMAE) for the parameter adaptation is introduced. In ASFM, measurements of outflow level are required every sample time. Applying ASFM to two catchment datasets, namely for the Pyungchang River in 1988 and the Chungju Basin in 1995, we show that the proposed ASFM has better performance than the conventional SFM.

抄録全体を表示

This paper presents a method for searching for the optimal paths for autonomously moving agents in mazes by modified Learning Vector Quantization (LVQ) in a reinforcement learning framework. LVQ algorithm is faster than Q-learning algorithms because LVQ concentrates on the best behavior in available behaviors while Q-learning algorithms calculate values of all available behaviors and choose the best behavior among them. However, ordinary LVQ sometimes mis-learns in the reinforcement learning environment due to erroneous teacher signals. Here a new LVQ algorithm is proposed to overcome this problem, which finds the optimal path more efficiently.

抄録全体を表示

We propose a learning rule for a layered neural network, which is based on backpropagation though unsupervised. Instead of a supervisor, the network gets a reward or a penalty from the environment. One characteristic of the network is a stochastic nature of the output units, which enables the learning through a tentative target signal set from a given reward. Namely, the network tentatively assumes the present output correct if a positive reward is given, and incorrect if given a negative, and backpropagates the corresponding error correction as an ordinary backpropagation. The algorithm is especially effective for learning the series of actions, under the additional framework that the reward is not only used to evaluate the present output but also assigned to the temporal series of outputs which result in the present reward. Thus the basic algorithm is extended in two directions to be applied to the time series learning. Type 1: Temporally accumulated values of errors are saved for a future reward, and once a reward is given a chain of outputs are evaluated. Type 2: Neurons with temporal integration of the inputs. Both types 1 and 2 of the extension are applied to time series calculations and several complex motion acquisition tasks of an autonomous agent, which is a simulation model of Khepera. It successfully learns the obstacle avoidance and capturing foods in several different environments using the proposed methods. The entropy of the input patterns during the learning is proposed as an index of the complexity of the learning environment, aiming to characterize the generalization ability to an unlearned environment.

抄録全体を表示

A new Q-learning method for the cases where the states (conditions) and actions of systems are assumed to be continuous is proposed. The components of Q-tables are interpolated by fuzzy inference. The initial set of fuzzy rules consists of all the combinations of conditions and actions relevant to the problem. Each rule is then associated with a value by which the Q-value of a condition/action pair is estimated. The values are revised by the Q-learning algorithm so as to make the fuzzy rule system effective. Although this framework may require a huge number of the initial fuzzy rules, we will show that considerable reduction can be done by adopting what we call “Condition Reduced Fuzzy Rules (CRFRs)”. The antecedent parts of CRFRs consist of all the actions and appropriately selected conditions, and their consequents are set to be their Q-values. Finally, experimental results show that controllers with CRFRs perform equally well compared to the system with the most detailed fuzzy control rules, while the total number of parameters that have to be revised through the whole learning process is considerably reduced with increasing the number of revised parameters at each learning step.

抄録全体を表示

We have investigated the method to measure the elastic moduli of composite materials by impact sound. For estimating the degree of fatigue of the materials, it is important to detect long-term components as well as periodic components from the impact sound waveform. In this report, we propose to use answer-in-weights neural networks for measuring the damping ratio of actual sound waveform.

抄録全体を表示

In this paper, a system identification method for linear regression models based on support vector machine is proposed. It is shown through a numerical example that the proposed identification method is robust for input-output data with outlier.

抄録全体を表示

We discuss how differences in personality, not functional differences, among members of a group affect its cooperative work. From the results of computer simulation of robots, the heterogeneous group performs the task more effectively than the homogeneous group.

抄録全体を表示

We have already reported that unbalanced input delivered to three-phase induction motor is useful to estimate rotation speed. In this paper, operation of control system with the estimated speed feedback is presented.

抄録全体を表示