As an engineering application of a biological immune system, an immune feedback control law featuring rapid response to foreign materials and quick stabilization of the immune system is applied to controlling the learning of multi-layer neural networks. Applying the immune feedback control law to the generalized δ-rule, the immune-feedback-based learning rule is derived. The stability condition of the proposed learning rule for the 2-layer neural networks is analytically described. To investigate the feasibility of the proposed learning rule to the multi-layer neural networks, simulation studies of identifying nonlinear functions using the 4-layer neural networks and the sandglass-type neural networks are carried out. Simulation results show both the effectiveness and characteristics of the immune feedback-based learning rule.

抄録全体を表示

The local image binarization is a useful method to convert a gray image to a binary image. In this method, a gray image is separated to some local areas and each area is binarized by the threshold determined in the area. However, the binarization is processed forcibly even if the brightness in an area is even. Therefore, the natural binary image cannot be always obtained by the visual sense because the discontinuity often happens between neighboring areas. This paper proposes a method to determine a set of thresholds for local image binarization using a genetic algorithm. The individual fitness in the population is defined by the matching rate of edges between the gray image and the resulted binary image. The experimental results show that the gray images were converted to the natural binary images by the proposed method.

抄録全体を表示

Automatic classification of bills has become important according to the progress of office automation. This paper is concerned with the new development of bill money classification based on a competitive learning algorithm where Italian Liras are adopted for classification. The Learning Vector Quantization (LVQ) method is used as a competitive learning. Original data of Lira bills may be rotated and/or shifted. We show that the LVQ method could be used effectively to classify the Lira bills under such various conditions.

抄録全体を表示

This paper proposes a method to classify new and used bills using the acoustic cepstrum pattern by neural networks. The proposed method deals with an acoustic signal which has been generated by the bill passing through a banking machine. By using an acoustic cepstrum pattern, the rough structure of the acoustic spectrum can be represented as the cepstrum with a smaller pattern size than the spectrum. The proposed method employs a neural network as the classifier. Two different types of the neural network, one is the three layered perceptron and the other is the competitive neural network, are used to evaluate which type is more suitable for classification of the cepstrum pattern. The experimental results show the effectiveness of the proposed method, and that the competitive neural network yields better classification performance than the three-layered perceptron.

抄録全体を表示

In this paper, we proposed a new hormonal-based Hopfield model neural network which is based on the idea of the homeostasis control of hormones. The network consists of a traditional Hopfield network and an evaluation network. The traditional Hopfield model network works depending on their initial states. We describe a conception of hormones to reset the initial states. A hormonal-based 4-bit Hopfield neural Analogto-Digital converter is designed and simulated by C language. Simulation shows that the local minima of the Hopfield model A/D converter can be eliminated by controlling the hormonal value. The hormonal-based Hopfield neural A/D converter performs A/D conversion much better than the traditional Hopfield model A/D converter. Furthermore, some interesting biological internal secretion phenomena are simulated and several experimentally testable predictions are made.

抄録全体を表示

This paper describes a learning multiple-valued logic (MVL) network that can explain reasoning. The learning MVL network is derived directly from a canonical realization of MVL functions and therefore its functional completeness is guaranteed. We develop traditional back-propagation to the MVL networks and drive a specific algorithm for the MVL networks. The algorithm combines back-propagation learning with other features of MVL networks, including the prior human knowledge on the MVL networks, for example, the architecture, the number of hidden units and layers, and many other useful parameters for the networks. The prior knowledge from the MVL canonical form can be used as initial parameters of the learning MVL network in its learning process. As a result, the prior knowledge can guide the back-propagation learning process to get started from a point in the parameter space that is not far from the optimal one, thus, back-propagation can fine-tune the prior knowledge for achieving a desired output easily. This cooperative relation between the prior knowledge and the back-propagation learning process is not always present in neural networks. The learning process in the MVL network also shares some cytology behaviors, in particular the cell adhesion, the cell aptopsis (the death of cell), and the cluster cell aptopsis (the death of cluster cells), and presents these properties in the artificial MVL network successfully. Simulation results are also given to confirm the effectiveness of the methods.

抄録全体を表示

We observed the coseismic electromagnetic phenomena at the occurrence of the earthquake in Sendia City area. This earthquake's epicenter was very close, less than 10km from our observation site. Our site accidentally located on the fault of which the underground tip slid at the focus. In addition to variations of the earth potential difference and the ground acceleration signals, we observed the clear variations of the ground ion concentration and the geomagnetic field signals. We cannot yet explain the cause how these variations appear and what conditions are needed to observe always these variations. More events are needed to make this clear and the observation must be continued.

抄録全体を表示

We have developed micro actuators using ion drag force generated by applied DC voltage. Accelerating electrodes for positive and negative ion were alternately placed, and a half of each elec trode was covered with insulating film. The operating liquid was a mixture of silicone oil and butyl alcohol. Pressure was increased rapidly to the number of active pairs of electrode in the ion drag pump with such structure of 18mm in diameter and 30mm in total length. When a set of subsidiary electrodes covered with insulating film was used instead of covering a half of each electrode, the flow direction was reversed. This was explained by formation of an ion sheath around each subsidiary electrode, since the subsidiary electrode covered with insulating film prohibits charge exchange. The ion sheath would build up a new electric field opposite to that of the applied voltage. This effect was then applied to a novel ion drag motor, which had subsidiary electrodes between each of the main bare electrodes. In this new motor, by changing the polarity of the applied voltage to the subsidiary electrodes, we could successfully control the rotating direction of the motor.

抄録全体を表示

Reinforcement Learning is the problem faced by a controller that must learn behavior through trial and error interactions with a dynamic environment. The controller's goal is to maximize reward over time, by producing an effective mapping of states to actions called policy. To construct the model of such systems, in this paper, we present a generalized learning automaton approach.
Comparing to Q-learning, the computational experiments of the pursuit problems show that proposed reinforcement scheme obtains better results in terms of convergence speed and memory size.

抄録全体を表示

The final purpose of this study is to accomplish the non-invasive imaging system of temperature-change distribution in the human head. The human tissue can be generally expressed with electrical resistance and capacitance. Capacitance is related to permittivity with temperature characteristic. In our method, capacitance is measured by many small electrodes arranged in surface of the human head. Using their values, permittivity distribution is reconstructed, and then, the distribution is replaced to temperature-change distribution using the permittivity-temperature characteristic.
In our previous study, a new capacitance measurement method was proposed and inhomogeneous permittivity distribution of brain phantom inserted agar-agar was reconstructed using the regularization method based on the modified Newton-Raphson method.
In this paper, the reconstruction of temperature-change distribution in the brain phantom is demonstrated when temperature of brain phantom is actually changed. As results, although the improvement in the measurement system and reconstruction algorithm is needed, the pattern of temperature-change could be presented. Therefore, the realization of the imaging of the temperature-change distribution in the brain phantom by means of this proposed capacitance measurement is expected.

抄録全体を表示

Reverse conducting thyristors are analyzed using a two-dimensional simulator to investigate the effects of substrate types, P-type or N-type, on transient characteristics. The ramp currents with four kinds of rise-time are applied to the thyristors. The thyristor with a P-type substrate shows faster turn-on time and lower clamping-voltage change than that of an N-substrate for the applied ramp currents. The excellent transient response to the P-type substrate thyristor is caused by the accelerated carrier, which results from the high electric field appearing in the P-base region.

抄録全体を表示

The deterministic annealing (DA) method for vector quantization can obtain better representative vectors than the LBG, CL, FSCL and SOFM. But it has two problems. One is that quality of the results is sensitive to the reduction of parameter T which is a temperature in an annealing operation. Another is that learning time is relatively long. Recently, the equidistortion principle has been derived and it is shown effective for vector quantization. In this paper, to solve above problems, a new algoritym named DAEP (Deterministic Annealing method with Equidistortion Principle) is proposed. In the DAEP, the idea of equidistortion principle is introduced into the DA method. In the annealing operation, location of representative vectors are forced to move to the global optimum points using the equidistortion principle. Judging from quality of results and learning time, it is shown that the DAEP can obtain better results than the other methods.

抄録全体を表示

For an accurate on-line measurement of ship's attitude the authors previously developed a measurement system which used one servo-type accelerometer and two servo-type inclinometers which were appropriately located on the ship. The system introduced, as dynamic models of the attitude signals, linear dynamic systems characterized by the candidates of the angular frequencies on the signals and used a bank of Kalman filters to realize an on-line measurement of the signals and also to overcome possible changes in the parameters as time lapse. Although the method realized an on-line accurate measurement of ship's attitude, it still needed to select adequate candidates for the parameter on a trial and error. This paper discusses the optimal location of the candidates based on a distance analysis in signal space and shows the effectiveness of the proposed approach by simulations.

抄録全体を表示

A new block adaptive algorithm which gives analytically an approximate optimum step gain in the Block LMS-Newton algorithm is presented. We further propose a method to reduce the amount of computations to O (N) computation per sample. Various simulations show that the proposed methods give fast convergence characteristics and reduced total computation time.

抄録全体を表示

In actual multivariable control systems, controlled variables interact each other. In this case one compensation for a control loop exerts harmful influence upon another control loop through the interaction. This paper presents a noninteracting PID control for a fluid temperature and liquid level interacting system. First, a precompensator is designed to remove the interaction of the multivariable control system using the inverse Nyquist array method. Further, the integrator decoupled system with state feedback is designed in the time domain. Then the design of the control system for the interacting system results in the individual compensation for the single loop PID control system. The experimental results showed the effectiveness of the precompensator designed by the inverse Nyquist array and the integrator decoupled system for removing the interaction.

抄録全体を表示

Quadric analysis of frequency turn-over temperature characteristics vs. orientations of quartz crystal resonators in the vicinity of circular AT-cut plate was carried out in the temperature range from -196 to 250°C, and these characteristics were discussed in comparison with the experimental results.

抄録全体を表示

We propose a new mutation operator with an adaptively variable parameter selection probability in the genetic algorithm. The genetic algorithm is applied to optimize the fuzzy reasoning in this paper. Optimization of the fuzzy reasoning consists of two kinds of problem. These are a combinatorial and a numerical optimization problems which are found in optimizing configuration of fuzzy rules and optimizing shapes of membership functions of the fuzzy reasoning respectivelly. Although the genetic algorithm can be simultaneously applied to both two kinds of problem, the genetic algorithm has not applied to optimize the shape of the membership functions in many cases, because the genetic algorithm can not quickly obtain a solution which can be satisfied. To quickly obtain such a solution using the genetic algorithm, we have already proposed a variable bit-selection probability in the mutation operator. However, parameter selection probability, one of factor to configure the genetic algorithm, also greatly affects result of the optimization. To improve this difficulty, we propose the adaptively variable parameter selection probability which serves to a rapid convergence to the optimized solutions.

抄録全体を表示

In this paper, we propose a dynamic cost selection method in the minimum cost flow optimization for wide-area water supply networks. We formulate the water flow on the multi-layered networks, considering the storage of all the reservoirs. Piecewise linear cost functions are employed so that the cost function can represent open and closed pipelines flexibly. With this characteristic of the piecewise linear cost function, various kinds of networks can be represented without changing the original network model. The optimization is executed under the selected set of cost functions. The simulated plans under the real data in a certain city show the advantages of the proposed method, in terms of confirming the operating conditions and saving water, even in emergency as well as the normal situations.

抄録全体を表示

This paper argued the number of VCV instances to be recorded in the synthesis unit dictionary as well as selection methods for small scale speech synthesis system. We proposed two selection methods, i) selection method by using phonemic environmental resemblance score (PER method), ii) selection method by searching minimal LSP distance path (MLD method). We evaluated synthesized speech generated by these methods. The followings were clarified, a) improvement in quality of synthesized speech was saturated with 14, 000 VCV instances. b) There was scarcely any difference in quality of synthesized speech between PER method and MLD method through the subjective comparison test.

抄録全体を表示

Nowadays it is not enough to supply products based on customer's preference, and both manufacturers and distributors must have excellent order fulfillment. Supply chain management (SCM) is a concept to facilitate effective product/cash flow control in global environment, and improving SCM is very important for increasing competitive position and profitability. Product allocation management in SCM is conventionally based on simple Theory of Constraints (TOC) concept, but it is not always provide pareto-optimal solutions. Market mechanism solves the product distribution problem by allocating the scheduled resources according to market prices. We formulate SCM system models as a discrete resource allocation problem, and demonstrate the applicability of market economics analysis to this framework. The proposed algorithm facilitates sophisticated product allocation system, which conducts pareto-optimal solution in the resource allocation problem.

抄録全体を表示

An attempt is made to improve well-known sufficient Schur stability conditions for real polynomials. The existing conditions are easy-to-check but conservative and used therefore as a preliminary test before resorting to exact ones. Pivotal to the scheme is an equivalence relation between Schur property of a polynomial and that of another one including a single free parameter. The resulting freedom enables to generalize the conditions and thereby lessen their conservatism. Numerical examples illustrate gains in sharpness attained by the proposed approach over existing conditions.

抄録全体を表示