We investigate a reinforcement learning of walking behavior for a four-legged robot. The robot has two servo motors per leg, so this problem has eight-dimensional continuous state/action space. We present an action selection scheme for actor-critic algorithms, in which the actor selects a continuous action from its bounded action space by using the normal distribution. The experimental results show the robot successfully learns to walk in practical learning steps.
We have already proposed the “rings gymnastic robot” that has free-floating characteristics of gripping point. The purpose of this paper is to acquire performance skill represented by fuzzy rules to realize one of the compulsory requirements, i.e., a backward giant circle from handstand to handstand, in the rings event through the robot. In controlling the performance, it is divided into three basic exercises and one of fuzzy controllers for the basic exercises is selected according to the situation. In order to acquire parameters of the controllers, an evaluation method based on five shbgroups is applied to genetic algorithm. A subgroup has independent evaluation and priority exists among the subgroups. Simulation results show the effectiveness of the performance skill obtained and the evaluation method proposed.
Recently, systems are becoming more complex and larger than ever, so numerous attempts have been made by researchers to introduce biological features into artificial systems, because many biological systems in the nature exist as one of the most complex systems. In this paper, we paid attention to symbiotic learning and evolution among creatures. The purpose of this paper is to introduce a new modeling technology using the concept of symbiosis, and propose a new learning and evolutionary method named Masbiole for the multiagents systems, which uses symbiotic strategies such sa mutualism, competition, predation and altuism to realize the symbiotic relations among agents consisted of neural networks. Further more, it is stated that Masbiole has pareto optimal solutions instead of common optimal solutions. From the simulations of Masbiole, it has been clarified that Masbiole can be a basic algorithm to realize the multi agent systems with symbiotic relations.
Many methods that can automaticlly make the behavior sequences of agents have been reported. Recently, a new evolutionary computation method named Genetic Network Programming (GNP) has been proposed. GNP is an extension of Genetic Algorithm (GA) and Genetic Programming (GP). GNP has a net work structure and it can make the behavior sequences of agents effectively. In this paper, an online learning method for GNP is proposed. The method can make GNP adapt to the dynamical environments by changing its behavior sequences immediately after GNP causes bad results.
This paper applies a method, Genetic Algorithm with Search Area Adaptation (GSA), to the function optimization. In Previous study, GSA has proposed for the floorplan design problem and it has shown better performance than several existing methods. We believe that investigation of the searching behavior of the algorithm is important. However, since the floorplan design problem is combinatorial optimization problem, we do not know in detail why GSA works well. To study details of the searching behavior, we believe GSA should be applied to a problem in which several benchmarks whose optima and landscapes are known have been proposed and an appropriate measure of distance between solutions can be defined. In this paper, we apply GSA to the function optimization, which satisfies the requirements. There is another purpose to apply GSA to the function optimization. We would like to propose a superior method for the function optimization. through several experiments, we have confirmed that GSA works adaptively along with the searching stage and it shows higher and stabler performance than one of existing methods.
We propose an adaptive probability density function (PDF) to select an effective action on reinforcement learning (RL). The uniform distribution function and the normal distribution function of an action are often used to select an action. When these fuctions are used, however, the information of search direction is net considered. The proposed method utilizing the information of it enables RL to reduce the number of trials, which is needed to real environment learning. Furthermore, the proposed method can be applied easily to various methods of RL, for example, actor-critic, stochastic gradient ascent method. The performance of our proposed method is demonstrated by computer simulations.
Neural network hardware is necessary for demanding real-time applications such as pattern recognition. Over one million gates of the latest FPGA (Field Programmable Gate Array) are available for the dedicated neuro hardware. In this paper, an improved calculation algorithm of the neuron model with sigmoid function is proposed, which is suitable for hardware implementation. The proposed algorithm is based on the multi dimensional binary search. The neuron circuis implemented on FPGA by the proposed algorithm have shown the excellence in size and circuit frequency compared with the conventional circuits with sum of product operation.
An integrative two-stage optimizaition method based on genetic algorithms (GA) using an if-then heuristic rule was developed to generate optimized boiling water reactor (BWR) loading patterns (LP), In the first stage, the LP is optimized using an improved GA operator. In the second stage, an exposure-dependent control rod pattern (CRP) is sought using GA with an if-then heuristic rule. This algorithm provides both good convergence performance and global searching ability. Therefore this optimization system realized the practical application for the real BWR LR design which does not need a reference LP and CRP. Through Calculation in an actual plant, it was confirmed that the optimization could be realized within a reasonable computation time. By automation and improvements in efficiency, the aim of automating a series of reload core design tasks of the BWR was accomplished. A considerable reduction of the work in the reload core design task can be expected.
Adaptation to dynamic environments is an important application of genetic algorithms (GAs). However, there are many difficulties to apply the GA to dynamic environments. Especially, in online environments, the GA's defects become remarkable because individuals should be evaluated in the real world. In this paper, we proposes a novel approach to such an online adaptation called the Environment Identifying Genetic Algorithm (EIGA). The EIGA achieves the online adaptation and identification of the environments simultaneously by the parallel technique and reduce the number of fitness evaluations in the real world by utilizing the identified environment. The thermodynamical selection rule is also utilized to maintain diversity. The performance and the adaptation ability of proposed method are confirmed by computer simulation taking a changing Nk-landscape model as an example.
This paper deals with modified scheduling problems in manufacturing systems. Recently optimization is required for not only basic scheduling problems but also scheduling problems modified in such a way that some constraints are added or are removed. Genetic algorithm (GA) is considered to be a promising method for solving the modified scheduling problems, because change of the procedure is small for the modification. In GA the problems can be solved only by adding (removing) the modules related to the added (removed) constraints. In this paper, a module type GA (MTGA) is proposed based on this idea, and the MTGA is designed for solving a mixed problem of machine scheduling and worker allocation. In a factory, both skillful workers and untrained workers work together, and the efficiency of operations depends on the allocation of workers to machines. The effectiveness of the proposed algorithm is investigated through a numerical result for a job-shop scheduling problem.
This paper addresses electric equipments configuration problems in a electric power plants. These equipments include ventilators, conveyance devices, electromotive valves, and so on. The problems are intrinsically large-scale combinatorial optimization problems with various constraints. The objective of the problems are to minimize both total cable lengths and their operating costs. These objectives may be contradictious. Furthermore, we would like to have multiple candidate solutions. Thus, to solve them, we must develop novel evolutionary algorithms. In this paper, we apply Bayesian Optimization Algorithm with Tabu search (Tabu-BOA) to the problems. From intensive experiments, the proposed method has shown the effectiveness to practical problems.
In this paper, new linearization technique for bipolar OTAs using exponential-law circuits is described. The core circuit of the proposed OTAs is the multi-TANH doublet. Two kinds of the OTAs have been designed. One has parallel configuration and the other is adaptively biasing configuration. The core circuit is combined with the SINH circuit for the parallel configuration or the COSH circuit as an adaptively biasing current source. In addition, two other designs of linear OTAs are proposed, and thus altogether four new linear OTAs are presented. Although these designs of the proposed OTAs are different, the theoretical normalized transconductances of the OTAs are expressed as the same function. The performance of the OTAs is compared with that of the multi-TANH circuits, which are known as conventional linear bipolar OTAs. The linear input voltage ranges of the OTAs are wider than that of the multi-TANH doublet and almost the same as that of the multi-TANH triplet. Furthermore, the power dissipated in the proposed OTAs is lower than that in the multi-TANH doublet and triplet. SPICE simulation shows that each OTA has different advantages.
Aged people who live alone are in particular need of a daily health check, medication, and of warm communication with family and friends. Especially, a daily health check terminal such as a blood pressure and a coughing, is very important for primary care. To achieve the high speed service for a daily health check with a home computer, we propose a prototype model of a cough data collecting and classification system. In this paper, we focus on the classification technics of the cough data for life support system. First, we tried to discriminate between vowel and cough. Secondary, we discriminate between vowel and cough with white noise. Under 70 dB vowel or white noise intensity to 90 dB cough intensity, the cough feature can be detected. Over 20 S/N ratio, it seems to be more practically reliable recognizing zone, which separate cough from noisy environment. For the analysis and processing of compound cough waveform, it is confirmed experimentally that a differential coefficient is one of more important features of cough.
A 1.3 GHz-pulse Doppler radar called a lower atmospheric wind profiler (LAWP) was developed by Communications Research Laboratory in 1990. The LAWP can measure lower atmospheric winds between 100 m and 3 km above the ground. However, the influence of ground clutter on LAWP measurements severely degrades the reliavility of wind velocity estimation, because LAWP detects very weak signals from clear air. On the other hand, Li has proposed a complex notch filter (CNF) by which sinusoidal signals in a quadrature sampling system can be detected. This filter is suitable for detecting complex signals, but it is very difficult to estimate an exact wind velocity by CNF, because of the influence of ground clutter and the fluctuating target. The purpose of this paper is to estimate the exact wind velocity under the above-mentioned ill conditions. The proposed signal processing consists of a complex notch filter and a complex adaptive line enhancer. Passing the received signal through the notch filter, the target signal is obtained. Then, passing this signal through the complex adaptive line enhancer, the wind velocity will be exactly estimated. The validity of this method is demonstrated by real data provided by LAWP.
Blind deconvolution is a method of recovering transmitted signals from only received signals. The minimum entropy method is one of blind deconvolution methods. This method has two problems that has slower convergence, and that its reliability is lower. In this paper, we propose the new algorithm for solving two problems above. The proposed algorithm is as follows. 1) it is based on the block processing with finite recovered signals. 2) Kurtosis is estimated by recovered signals with each block. 3) Cost function is decided by kurtosis. 4) Transmitted signals are recovered by received signals using decided cost function during the block. We confirm validity of the new algorithm by computer simulation.
In this paper we present a new technique to restore the 3-D facial shape using the stereo camera and the pseudo-random colored slit ray pattern. Because of the flatness of the forehead and cheeks and changeless complexion on a face, it is rather difficult to find the corresponding points on the left and right stereographic facial images. To overcome this difficulty, we project red and green slit ray pattern on a face by an overhead projector. From the coordinates of the corresponding colored slits on the left and right images, we acquire the depth information and then restore the facial shape. To verify the exactness of the restored facial shape, we rotate the restored face and extracted the contour line of the profile and compared with the profile of the same subject. The result shows a remarkable coincidence of two images and this proves the usefulness of the proposed technique.
In this paper, an approach to feature extraction utilizing independent component analysis (ICA) is pro-posed. In our approach, input patterns are transformed into feature vectors using ICA-bases that are obtained through two-layer neural network learning. A k-NN classifier is applied to these ICA feature vectors when the recognition accuracy is evaluated. Hand-written digits in MNIST database are used as target characters. Fast ICA algorithm is applied to these images in order to learn ICA-bases. In recognition experiments, we demonstrate that the ICA approach realizes a potential feature extraction method for hand-written digits. Furthermore, we show the addition of noise patterns to training data is effective for elimination of redundant basis functions.
Several researchers have recently attempted automatic recognition and tracking of the facial expressions in video frames. However, it is very hard to extract facial movement because the distinction between facial movement and head movement is difficult to make. It is necessary to make the initialization of the head position using the central points of each eye and the tip of the nose in order to solve this distinction problem. The tracking of facial movements is made in comparison between the neutral face and the expressive face. The detection of neutral face is done with the condition, “facial movement is very small within the limited interval”. However, some expressive face were also detected as a neutral face. Therefore, the function to evaluate the detected neutral face is adopted to correct the false detection of a neutral face. The transition of facial expressions can be recognized by means of the detected reference face (neutral face) and the recognition algorithm constructed to the fuzzy expert system with flow of application of rules.
There are lots of researches on designing self-tuning control(STC) systems for linear nonminimum-phase systems. A wide variety of existing methods on designing STC systems are conceptually classified into the following two groups. The one is based on minimization of an error criterion, and the other is on pole-assignment of the closed-loop system. On the other hand, in order to guarantee the tracking property for the reference signals, the internal model which corresponds to the reference signals, has to be inserted into the control system based on the internal model principle. However, the variance of the control error signal may become large by inserting the internal model, because the stochastic noise may be amplified. In this paper, a design method of self-tuning pole-assignment control systems with internal models is proposed based on minimizing the linear quadratic cost function. The newly proposed control scheme has the feature that the variance of the stochastic noise in the control system is reduced owing to the minimum variance control scheme, and also the tracking property to the reference signal is assured owing to the internal model. The proposed control scheme is numerically evaluated on some simulation examples.
An application of a new multi-scale stochastic theory called the Multiple Tree (MT) theory, will be presented in this paper. The MT construction itself is an aggregation of several single trees where these trees are connected each other through a Gaussian random vector. The nature of the MT is suitable to the computer network environment. In this paper such advantage will be exploited in estimating 1/f signal given ill-posed observation over the network.
This study is concerned with detection of objects in motion from time serial pictures of a 6-D. of freedom motion monocular camera. As a point in real space throw an image on corresponding pixels of time serial pictures. So to solve this problem to find out the corresponding algorithm is essential. In this paper we proposed a new method of the detection of motion objects based on compensating background image shift caused by camera rotation and translation. The main point of this idea is, to take two pictures by the passive sensor (CCD camera) in two different times tk and tk+_??_ while the platform is moving, by correcting them rotationally (also called vertical rectification) the rotation disparity will be excluded, and we get two parallel frames that will be matched through their histograms for the translation-distortion extraction.
In this paper, we propose a new system for recognition of scenery images. The proposed system consists of 2 part. One is the part which infers based on absolute position of the target. The other is the part which infers based on relative position of the target. Each part consist of Fuzzy Inference Neural Network (FINN)(10) which can extract fuzzy if-then rules automatically. So the system can extract knowledges about absolute position and relative position. Through computer experiments, it can be seen that the proposed system can recognize the image much correctly.
This paper describes a system that confirms vehicle drivers' safety based on the images taken by the camera located in front of the vehicle before going into the intersection. While vehicle approaches an intersection slowly, this system detects another vehicles approaching to the same intersection. We accomplish the detection by removing background regions using frame-difference. In order to extract vehicle regions correctly, it is necessary to adjust the interval of frame-difference according to the location of the vehicle extracted from the previous frame. Furthermore, in order to extract both farther vehicle and closer vehicle correctly, two kinds of interval, shorter and longer, are adopted. Also, when the vehicle is running, the background regions of taken images change non-lineally. Therefore, it is necessary to transform images taken from different viewpoints to images observed from one same viewpoint by perspective transformation. Based on the transformed images of each frame, the vehicle regions can be extracted using frame-difference. Moreover the running directions of extracted vehicles can be determined by tracking the vehicles' location for several frames. Experimental results show that the proposed method is efficient.
In this paper, an FPGA (Field Programmable Gate Array)-implementable chaos circuit with array structure is proposed. The array structure which consists of 1-dimensional chaos circuits enables us to construct an S-dimensional chaos circuit (S=1, 2, 3, ...). Furthermore, the circuit possess exact reproducibility of output signals. The validity of circuit algorithm is confirmed by numerical simulations. The proposed circuit is designed by Verilog-HDL (Hardware Description Language). The experiment concerning the HDL designed circuit shows that the proposed circuit can be implemented into the form of an FPGA.
Human footstep recognition (walk-recognition) in two wooden houses is discussed by comparing the spectrum envelope of footsteps. As a feature parameter, the correlation coefficient between footstep spectra is used. A simple walk-recognition algorithm is implemented. The average recognition rate is approximately 93%, and then the feasibility of distinguishing a walker and his footwear is confirmed.