In this paper, we examine the behavior of Evolutionary Computations under dynamic environments. In order to generate the dynamic environments systematically, we define the framework of the Dynamic Constraint Satisfaction Problems. Our comparisons of ECs to investigate the applicability to solve the DCSPs are based upon two viewpoints: the difference of reproduction manner and the use of local search methods, i.e., repair methods in the case of solving CSPs. The experimental results carried out in this paper show that Estimation of Distribution Algorithms with repair method could track the changed environments well.

View full abstract

N-gram indexing method is an algorithm for the full text search system where each index consists of serial N words or characters. To save the volumes of index file, the system has 2-gram index as base and to improve the performance, it has higher-gram index as supplement. This paper describes the performance of the incremental 3 gram index for the first step. Simulation studies are now on going.

View full abstract

We consider a method to recommend news articles by personal interests with a genetic algorithm that is employed for making the proper vector of user-profile. The user-profile is revised consecutively in order to keep the amount of calculation. The effectiveness of the proposed method is evaluated by simulation.

View full abstract

An Ontology is very important to read a technical document, but it costs too much to construct an Ontology. In this paper, we examined the advanced idea about the lexical extraction from a document, and lexical system creation, and evaluated it using the prototype system.

View full abstract

fMRI was used to investigate the neural substrates responsible for Japanese string reading. The subjects performed Sr (read strings from right to left) and Sl (read strings from left to right) tasks. According to the analysis results, a significant increase blood flow during the Sr task relative to Sl task was detected in multiple brain regions, including the lingual gyrus, fusiform gyrus, superior inferior and middle occipital gyrus, inferior temporal gyrus, superior parietal lobule, inferior and middle frontal gyrus, parahippocampal gyrus and medial aspect of the frontal lobe. These results strongly supported the conclusions of past studies about word perception. In addition, we found the spatial perception and working memory were particular importance for the perception of string.

View full abstract

We suggest a system that synthesizes music suitable for emotional expressions of dances. This system divides a dance into a few seconds periods. Then, It extracts factors from dancer's motion, and transforms them into an emotional vector,then selects the most suitable music by comparing it with vectors expressing emotions of music. We made a prototype system and evaluated it in subjective questionnaires.

View full abstract

While personal computers and their peripheral units become smaller and lighter, user interfaces are also required portability in stead of carrying conventional keyboard to type characters. Some devices have been proposed and provided for this demand. However, they have restrictions for functionalities and circumstances to use. This paper proposes a virtual keyboard with fewer restrictions for general use. The virtual keyboard detects key strokes by monitoring the motions of the fingers by a small camera equipped on the user's wrist.

View full abstract

We regard the sequence of returns as outputs from a parametric compound source. Utilizing the fact that the coding rate of the source shows the amount of information about the return, we describe l-learning algorithms based on the predictive coding idea for estimating an expected information gain concerning future information. Using the information gain, we propose the ratio w of return loss to information gain as a new criterion to be used in probabilistic action selection strategies. In experimental results, we found our w-based strategy performs well compared with the conventional Q-based strategy.

View full abstract

In reinforcement learning problems, agents should learn from only rewards that are provided by the environment; hence, learning by trial and error is inevitable. In order to acquire right policies of actions, action-value functions are often estimated. In many cases, the action-value functions are approximated by parametric linear/nonlinear functions such as RBF networks. However, when the RBF networks are trained in incremental fashion, we often suffer from a serious problem called interference that results in the forgetting of input-output relations acquired in the past. In this work, we propose a new approach to learning action-value functions using an RBF network with memory mechanism. In the simulations, we verify that our proposed model can acquire the proper policies even in difficult situations.

View full abstract

In this paper, an application of reinforcement learning to a radio-controlled helicopter is considered. The Actor-Critic algorithm is employed, where an Actor and a Critic are realized by using RBF neural networks. In the Critic, connection weights between neurons are updated by adopting the backpropagation algorithm to minimize squared TD-error. Control signals are generated by adding noises to the output of the Actor while learning. As for the Actor, connection weights between neurons are updated based on the evaluation of control results. Moreover, through some computer simulations using 2-dimensional model, it is observed that the network learns hovering flight controls.

View full abstract

In this paper, we apply reinforcement learning to acquire gait patterns of quadruped walking robot, AIBO, produced by Sony. The purposes of our study are to investigate a method of applying reinforcement learning effectively and to examine the applicability of reinforcement learning to a commercial robot, where the function are restricted to some extent. In the paper, we adopt the Actor-Critic algorithm with eligibility trace, and newly design a reinforcement learning method. Through several experiments by using our proposed method, it is observed that one of the typical gait pattern of quadruped can be acquired.

View full abstract

We introduce a new technique for the texture image modeling by using 2-dimensional periodic model. In case of using 2-dimensional autoregressive (2-D AR) model, a texture image with unstationary characteristics has difficulties in extracting accurate texture feature. We develop 2-dimensional periodic model to adopt 2-D AR model for the accurate feature extraction on such a texture image and show the validity to the modeling of the unstationary texture image.

View full abstract

Among the multi-resolution image processing methods, the pyramid linking algorithm has been reported to be useful in a large number of research fields. However, it is difficult to measure the similarity between child and parents in the pyramidal structure. In this paper, we propose a image segmentation method by optimizing pyramidal structure with genetic algorithms. The proposed method can divide an image composed of different kinds of texture fields into homogeneous areas without a function to discriminate similarity between child and parents.

View full abstract

Measuring the proportion of white blood cells is effective for estimating the source of the disease or for observing the progress of the disease. However, it takes long time by hand and requires the specialized knowledge, thereby the method for automatically classifying cells is needed. We propose a classification of white blood cells based on neural network in this paper. Nine parameters of area, shape and color information are extracted from a microscopic image of cells. These parameters are used as inputs for neural network classifier. Experimental results show that our proposed method can classify cells with high accuracy.

View full abstract

Applications of independent component analysis (ICA) to feature extraction have been a topic of research interest. Here, we propose a novel recognition method using features extracted by ICA. The proposed method consists of some modules for each category and a synthesizer. A module has a feature extraction and a classification. Features are independent components extracted by ICA algorithm using the training data for each class and classification are made by these features. These output of the module are combined and categories are decided by a majority rule. We evaluate the performance of the proposed method for several recognition tasks. From these results, we confirm the effectiveness of the recognition method using independent components for each class.

View full abstract

In this paper, we introduce a license plate recognition system based on a new image retrieval method. In this system, a plate image is dealt with a pattern, and the significant image feature of the pattern is used for the license plate retrieval, so that it is not necessary to recognize the each digit area and to distinguish each digit by using neural networks and/or fuzzy systems in traditional systems. The significant multiresolution information on the wavelet decomposed plate images are stored in advance as a database, and most similar image to the query is retrieved. Our system has an advantage to permit a position gap and/or a pattern fluctuation of the license plate by utilizing the multiresolution information of wavelet decomposition.

View full abstract

The purpose of this research is the recognition of road signs by a neural network.In this paper, we propose to use HSL color system or the double complementary colors for the inputs to ColorNN, which detects the color of road signs. The performances are evaluated by ROC analysis and compared with the previous result obtained by the NN inputs based on RGB color system. The experimental results show that HSL color system improves the detection of the red and yellow, while frequent misdetections of the sky area cause a low detection rate for the blue.On the other hand, inputs based on the double complementary colors give an improvement also in the detection of the blue.

View full abstract

Recently, information terminals such as mobile phone are widely used. However, these devices are not always comfortable and convenient to use because of their small shape. Consequently, we need to develop a new type of the input interface whose operation is easier and shape is smaller. Electromyogram (EMG) is generated along to the person's behavior and it has information, which contains of the level of behavior. Especially, learning and recognition of EMG by using neural networks, is supposed to be possible. We have shown its basic ability and possibility by the simulation for the realization of the practical use on the wrist behavior recognition. Furthermore, we consider about online tuning of neural network for robust of fluctuation of individual wrist behavior

View full abstract

An incremental or iterative learning classifier based on support vector machines (SVMs) is proposed and analyzed from the geometrical viewpoint. We show that the effective examples are necessary and sufficient to obtain the SVM solution of given examples. This means that SVM can be solved online. The proposed method stores support vectors instead of effective examples since it is difficult to discriminate effective examples. Some convergence properties of the proposed method are also given.

View full abstract

In this paper, we compare L1 and L2 support vector machines from the standpoint of training time and the generalization ability. For blood cell data, the generalization ability of L2-SVMs is a little higher than that of L1-SVMs but training time of L1-SVMs is usually shorter than that of L2-SVMs. We also compare the effect of the approximate KKT (Karush-Kuhn-Tucker) conditions using the bias term and the exact KKT conditions. According to the computer experiments, since the approximate KKT conditions give a conservative estimate of violating variables, training time using the approximate KKT conditions is usually shorter.

View full abstract

In this paper, we discuss fuzzy least squares support vector machines that resolve unclassifiable regions for multiclass problems. We define a membership function in the direction perpendicular to the optimal separating hyperplane that separates a pair of classes. Using the minimum or average operation for these membership functions, we define a membership function for each class. Using the blood cell data, we show that recognition performance of fuzzy LS-SVMs with the minimum operator is superior to that of fuzzy SVMs. While, although the performance of fuzzy LS-SVMs with average operator is inferior to that of fuzzy LS-SVMs with minimum operator, it is comparable to that of fuzzy SVMs.

View full abstract

Fuzzy c-Varieties (FCV) is the linear fuzzy clustering algorithm that estimates local principal component vectors as the vectors spanning prototypes of clusters. However, least squares techniques often fail to account for outliers, which are common in real applications. In this paper, we propose a technique for making the FCV algorithm robust to intra-sample outliers. The objective function based on the lower rank approximation of the data matrix is minimized by a robust M-estimation algorithm that is similar to FCM-type iterative procedures.

View full abstract

This paper proposes an ensemble learning method of fuzzy rule-based classification systems. Two different types of fuzzy rule-based systems are involved in our ensemble learning method. One is a fuzzy classification system whose output is the suggested class of an input pattern. The other type of fuzzy rule-based system assigns a credit to each fuzzy classification system. A boosting method is used to construct a collection of fuzzy classification systems. In the boosting method, a weight is assigned to each training pattern. Training patterns with a large weight are used to construct a fuzzy classification system. We show the effectiveness of the proposed method in computer simulations.

View full abstract

In this paper we propose the kernel method to calculate a Mahalanobis distance in the feature space and propose a kernel fuzzy classifier with ellipsoidal regions in that space. In our proposed kernel method, we first select linearly independent vectors that form a basis of the subspace in the feature space and discard remaining vectors. By this method, the covariance matrix in the feature space is not singular, and we do not need to use singular value decomposition. Thus we achieve training speed-up compared with the conventional kernel method. We evaluate our method using blood cell data. The result shows that the generalization ability is better than that of the conventional fuzzy classifier, which is generated in the input space. In addition, we can confirm that our proposed kernel method is effective for training speed-up.

View full abstract

This paper shows how a small number of fuzzy rules can be selected for designing interpretable fuzzy rule-based classification systems. Our approach consists of two phases: candidate rule generation by data mining criteria and rule selection by evolutionary multiobjective optimization (EMO) algorithms. First a large number of candidate rules are generated and prescreened using two rule evaluation criteria in data mining. Next a small number of fuzzy rules are selected from candidate rules using multiobjective genetic algorithms. Our rule selection is formulated as a combinatorial optimization problem with three objectives: To maximize the classification accuracy, to minimize the number of selected rules, and to minimize the total rule length. Thus the task of genetic algorithms is to find non-dominated rule sets with respect to the three objectives.

View full abstract

From the current research by using the algorithm based on Grouping Genetic Algorithm(GGA), it was proven to be effective in organizing the academic conference program. However, we have not investigated the content of the program of SCI'01 by this algorithm. In this observation then, we aim to review the created schedule of the conference in more detail. The result of this research will validate the effectiveness of the proposed algorithm.

View full abstract

The existing Genetic Network Programming (GNP) has a problem that the optimal number of nodes in GNP must be determined by trial and error. In this paper, we propose a method for determining the optimal number of nodes in GNP automatically, using genetic algorithm based on the coexistence of heterogeneous populations. Furthermore, we apply the proposal method to the food capture problem of ants, and demonstrate its effectiveness.

View full abstract

As an incremental learning model, we have proposed Resource Allocating Network with Long Term Memory (RAN-LTM). In RAN-LTM, not only a new training sample but also some memory items stored in Long-Term Memory are trained based on a gradient descent method. The gradient descent method is generally slow and tends to be fallen into local minima. To improve these problems, we propose a fast incremental learning of RAN-LTM based on the linear method. In this algorithm, centers of basis functions are not trained but selected based on the output errors. A distinctive feature of the proposed model is that this model dose not need so much memory capacity. To evaluate the performance of our proposed model, we apply it to some function approximation problems. From the experimental results, it is verified that the proposed model can learn fast and accurately unless incremental learning is conducted over a long period of time.

View full abstract

This paper deals with a determination method of exploring points for multi robot exploration. In this method, each robot can share exploring information of itself and other robots by the belief measures, and determine exploring points considering the belief measures. The effectiveness of our approach is demonstrated by a real experiment with two mobile robots.

View full abstract

This paper presents a method for human-computer cooperative control of a mobile robot. As long as a priori information given to the computer is accurate, the computer controls the mobile robot by following working commands programmed in advance. When it is necessary for the human operator to correct the motion of the robot because of some changes of the environment, the human operator intervenes in computer operation to reduce the speed of robot and change the course of locomotion by moving the joystick. Through some actual trials in a PWS mobile robot, the proposed method is proven to be efficient in the operation of a mobile robot with simplicities of commands and flexibility of human operators.

View full abstract

We have developed a seam tracking system that is attached to an industrial robot and achieves a tracking accuracy of 0.1 mm. The system consists of a weld line sensor and seam tracking mechanism. The weld line sensor detects two seam positions along a weld line simultaneously. Since the detected seam positions include robot position errors, the system then estimates the real seam positions and robot position errors by using two detected seam positions, and the tracking mechanism follows the real seam positions. The accuracy of the system has been examined by both computer simulation and experiment.

View full abstract

A knocker robot was developed. This robot can control the swing speed and the swing angle of a bat, and also the position of a ball. The ball speed just hit by the bat was measured by using a high speed camera. The trajectory of a ball can be simulated by a differential equation model for which the data from the high speed camera were used. The relationship between the swing speed and the initial ball speed, and also that between the initial ball speed and the flight distance were studied by experiments and simulation.

View full abstract

This paper applies the randomized alogorithm to the real μ analysis.It is shown that the computation time can be reduced considerably by using the alogorithm in some numerical examples,compared to the case where an existing LMI solver is used.

View full abstract

To design gain scheduled controllers in a less conservative fashion, a dilated-LMI-based approach has been proposed recently. However, this approach is still conservative due to the use of multi-affine parameter-dependent Lyapunov variables. In this paper, we revisit the dilated LMI condition and provide a yet less conservative design method that enables us to design gain scheduled controllers via polynomial-type parameter-dependent Lyapunov variables.

View full abstract