For sampling various solutions from the entire Pareto Front of the multi-objective resource division problem, a new Genetic Algorithm (GA) based on the evolutionary theory advocated by Kinji Imanishi is proposed. First of all, two types of distance between two individuals, namely, structural and functional distances, are introduced and used to define four types of relation between them, namely, homogeneous, heterogeneous, homologous and analogous species. Then, for keeping various species within the population as much as possible, a new generation model with variable population is presented. In order to search Pareto-optimal solutions effectively, a new genetic operation that combines the Harmonic crossover with a local optimization algorithm is also proposed. Finally, the advantage of the Imanishism-based GA is demonstrated through computational experiments conducted on two- and three-objective problem instances.
Recently, many methods of evolutionary computation such as Genetic Algorithm (GA) and Genetic Programming (GP) have been developed as a basic tool for modeling and optimizing the complex systems. Generally speaking, GA has the genome of string structure, while the genome in GP is the tree structure. Therefore, GP is suitable to construct the complicated programs, which can be applied to many real world problems. But, GP is sometimes difficult to search for a solution because of its bloat and introns and also because the effect of crossover and mutation deffers depending on which nodes are operated by crossover and mutation, therefore, sometimes premature convergences emerge in GP.In this paper, a new evolutionary method named Genetic Network Programming (GNP), whose genome is a network structure is proposed to overcome the low searching efficiency of GP and is applied to the problem on evolution of behaviors of ants in order to study the effectiveness of GNP. In addition, the comparison of the performances between GNP and GP is carried out in simulations on ants behaviors.
While searching for suboptimal solutions for large-scale problems, it is critical to force search algorithms on promising regions. This paper presents genetic algorithms with search space reductions (RGAs) and their application to solving large-scale permutation fiowshop problems. The reduced search spaces are defined by adding precedence constraints generated by heuristic rules. To balance between the size of reduced spaces and the risk of missing good solutions, a set of consecutively included search spaces is proposed. RGAs are implemented and their performance is tested on a large-scale flowshop problem. Primary experiments show that the RGAs outperform the standard genetic algorithms greatly. Moreover, we propose an improved uniform crossover operator which preserves the precedence constraints to focus genetic search on the specified search spaces. It is shown from computational experiments that the mechanism of search space reductions works well with GAs and RGAs outperform standard genetic algorithms significantly.
To avoid a traffic accident, it is desired to detect possible collisions between a vehicle and obstacles at high speed. However, high resolution of an obstacle representation results in increase of collision detection time. In this paper, we propose a collision detection VLSI processor based on a hierarchical algorithm. Unless any possible collision is detected in a coarser representation of obstacles, there is no need to detect possible collisions in a finer representation. The VLSI processor consists of several content-addressable memories (CAMs) for parallel matching operation and processing elements (PEs) for parallel coordinate transforma-tion. When the utilized ratios of a CAM and PEs are 100%, minimization of collision detection time under an area constraint can be attributed to the area-time product minimization of a CAM and a PE. As a result, the highest performance is achieved by using a ROM-type CAM and a bit-serial pipelined PE.
This paper presents a new type recurrent neural network and its learning algorithm for nonlinear dynamics named “Velocity-Error Backpropagation (VEBP).” In VEBP, learning is performed by 2 steps: (a) the velocity vector field of reference trajectories is approxi-mated by feedforward neural network with bi-connection layers by backpropagating velocity errors directly. (b) recurrent neural network is constructed by adding integrators and output feedback loops to the trained feedforward neural network. VEBP has some advantages with conventional learning method for recurrent neural networks named “back-propagation through time (BPTT).” Effectiveness of the presented recurrent neural network and its learning algorithm is demonstrated by simulation results for some examples of nonlinear dynamics.
HMS (Holonic Manufacturing System) is a distributed manufacturing system that consists of autonomous and cooperative manufacturing elements called holons. Individual holons share roles so that they can exhibit their maximum capability through the cooperation among themselves without being constrained by centralized mechanisms. This paper proposes manufacturing holons in HMS and cooperation mechanisms among them. Although the mechanisms cover only small parts of manufacturing activities, variety of combinations that is provided by holons has shown the feasibility for achieving highly adaptive manufacturing systems. These are results of HMS project in IMS (Intelligent Manufacturing System) program.
Traditional econometric models assume a constant one-period forecast variance. To generalize this implausible assumption, a new class of stochastic processes called autoregressive conditional heteroskedasticity (ARCH) processes were introduced in Engle (1982). This type of model behavior has already proven useful in modeling several different economic phenomena. In those papers, maximum likelihood estimation of the linear regression model with ARCH error was discussed. However, it is well known that there exists the nonlinearity in economic time series by empirical research. Recently, the extended correlation least mean squares (ECLMS) algorithm has been proposed to solve the double-talk problem in the echo canceling system. The characteristic of ECLMS algorithm is to utilize the correlation functions of the input signal instead of the input signal itself, to process and find the parameters of system. Noise signal is separated from observed signal by ECLMS algorithm. Therefore, the estimation performance is considerably improved. Parameters in nonlinear time series are calculated by ECLMS algorithms. We demonstrate that it is feasible to esti-mate second-order Volterra model with ARCH error by ECLMS algorithm and some numerical examples are presented to illustrate that the proposed method can work well for noisy signal using computer simulation.
To build and apply framework is a practical way to improve reusability of software products and to realize effective system development. A well-designed framework depends on correct estimation of the requirements for software modification in its system domain. In this paper, we propose “Device-Control-Constraint Model” as a reference model to build effective frameworks in control system domain. We design the reference model by assuming that the main requirements of system modification are physical device interface, device composition, and control algorithm. By our reference model, frameworks are composed of three sorts of objects. They are Device objects, Control objects, and Constraint objects. Each object is designed to deal with different requirements of modification. We discuss the design of Device-Control-Constraint model and show an example of framework construction in air-conditioner control domain by using our model.
Arterial baroreflex functions most importantly as a blood pressure adjustment system. Arterial baroreflex is the function in which arterial baroreceptors monitor blood pressure, working to maintain blood pressure at a certain level by means of sympathetic and parasympathetic nerve activity. The arterial baroreflex is usually tested by intravenous injection of drugs causing an increase or decrease in blood pressure and by measuring the slope of the regression line of ECG R-R interval and blood pressure time series. This paper proposes different methods, not involving the injection of drugs. In the methods proposed in this paper, the autoregressive (AR) model was applied to R-R interval time series during rest, and systolic blood pressure time series. By applying impulses to the AR model prediction error, an impulse reaction with equivalent response to changes in blood pressure was obtained. Using this function, we measured the slope of the regression line of ECG R-R interval, and systolic blood pressure time series. Analysis of the results yielded a correlation coefficientt of 0.8 between method 4 and the drug injection method. Although the value obtained from method 4 was smaller than that by the drug injection methods, the value was stable. Therefore, it is suggested that the methods proposed in this paper are valid.
In this paper, we described the detection characteristics of a semiconductor ring laser (SRL) as an optical inertial rotation sensor. First, we tried to measure the beat frequency characteristics as a function of the rotation rate, changing the loop radius for investigating the loop radius dependence on the detection sensitivity. Experimental results showed that the beat frequency caused by Sagnac effect was proportional to the loop radius and the detection sensitivity was proportional to the loop radius. For example, the sensitivity was increased _??_1.1 times by changing the radius from 12.4cm to 13.6cm. This result coincided well with expected value of 1.10. Furthermore, the beat detection characteristics depended on the oscillation condition of the SRL. From these experimental results, we could verify the possibility of a mode-locked SRL as a gyroscope for detecting stable beat note.
Artificial Cellular Neural Network (ACNN) is a kind of Neural Networks whose connections of its each unit only exist between itself and its neighborhood units. ACNN has several merits that layered Neural Network and mutually connected Neural Network do not have. For example, a unit in the ACNN does not have to connect a far unit, so it is easy to integrate units. Besides, the amount of data that is required to represent ACNN is less than that of fully interconnected Neural Network. The former is in proportion to the number of total units of the ACNN, though the latter is in proportion to the square of the number of total units of the fully interconnected Neural Network. In this paper, we introduce “Chase Game” to demonstrate an ability of ACNN. There are two agents in the “Chase Game.” One is a chaser, and the other is a runaway. We programmed the runaway's escape algorithms. On the contrary, the chaser's action is determined by a three dimensional ACNN. We used Genetic Algorithm (GA) in optimizing three-dimensional ACNN that can capture the runaway in short turns. And we confirmed that the chaser showed intelligent pursuit actions.
This paper deals with modeling, uncertain structure and μ-synthesis of a magnetic suspension system. The dynamics of magnetic suspension systems are characterized by their instability and complexity of electro-magnets, and they should be robustly stabilized in spite of model uncertainties. First we derive a nominal design model of the plant under some assumption, then we investigate the gap between the real physical system and the obtained nominal design model. This gap has complex structure which is expressed by the structured uncertainties that includes linearization error, parametric uncertainties, and neglected dynamics. Then we set the interconnection structure which contains the above structurally represented uncertainties. Next we design a robust controller which achieves robust performance using the structured singular value μ. Finally, we evaluate the proposed interconnection structure and verify robustness and performance of the designed μ controller by several experiments.
Hough transform is known as a useful method to extract a global figure in an image through the voting process into the parameter space. Because Hough transform uses the information about the edge pixels in an image for the voting process, the accuracy of the direction of each edge pixel influences on the reliability of the voted result. This paper proposes a method to raise the reliability of the voting process in Hough transform by improving the accuracy of the direction of an edge pixel using the directional adaptive filter. The direction and the width of the directional filter are determined dynamically by the direction and the curvature of the edge pixel. Experimental results show that the voted result with the sharper maximum peaks are obtained by the proposed method in comparison with the method that uses the direction of the independent edge pixel and the method that uses the averaged direction of neighboring edge pixels.
This study uses genetic algorithms (GAs) for the system which generates a finite automaton automatically. This system can build an automaton by incorporating automaton construction algorithms (ACAs) into the GAs. ACAs are algorithms that construct a finite automaton on the basis of gene information. The per-formance of this system is realized by integration algorithms and minimization algorithms. The integration algorithms are applied when an automaton is made by combining several automata, each of which is created by divided requirements of the original automaton. The minimization algorithms are used to minimize the number of nodes of the automaton which GAs and the integration algorithms generate. The most suitable finite automaton is formed automatically by using these algorithms collectively. An automaton drawing pro-gram can show the finite automaton as a state transition diagram to confirm the result easily. By simulations, the effectiveness of this method was confirmed.
A fast image compression and reconstruction method is proposed, based on various types of fuzzy relational equations, i.e., max-t-norm, adjoint max-t-norm, min-s-norm, and adjoint min-s-norm composite fuzzy relational equation, where the quality of reconstructed images is also improved. The experiment using 20 images from a standard image database (SIDBA) confirms the decrease in the compression (≈ 1/300) and reconstruction (≈ 1/310) times, as well as in the root mean square error (≈ 1/2), compared with the results of Hirota & Pedrycz (1999). It is shown by the wavelet analysis that the overall appearance of the reconstructed images is comparable with those of the discrete cosine transform method.
This paper develops a tracking control problem of rigid-link electrically-driven robot manipulators with uncertainty. We extend a research line of Ishii et al. (2000) to the case where there exists uncertainty such as modeling error or parameter perturbation in the robot dynamics. First, torque level control input to achieve tracking control is applied for manipulator loop, and filtered tracking error system is derived. Then, penalty signal for tracking error between the outputs of the manipulator and desired trajectories is introduced and the tracking performance is evaluated by L2-gain from torque level disturbance signal to the penalty signal. Finally, two control laws, i.e. robust adaptive and modified robust adaptive tracking control laws with L2-gain disturbance attenuation, are designed such that the closed-loop error system is globally stable in the sense of uniform ultimate bounded stability with the L2-gain less than any given small level. Experimental works were carried out for a two-link electrically-driven manipulator. The results show the effectiveness of the proposed control scheme.
Data mining is part of a larger area of recent research in Artificial Intelligence and Information Processing and Management otherwise known as Knowledge Discovery in Database (KDD). The main aim here is to identify new information or knowledge from database in which the dimensionality or amount of data is so large that it is beyond human comprehension. Self-Organising Map is used to analyse power transformer database from one of the electric energy providers in Japan. Furthermore, the regression aspect of SOM is also tested. Regression is achieved by searching for the Best Matching Unit (BMU) using the known vector components.
The fuzzy control strategy of superconducting magnetic energy storages (SMESes) was proposed for leveling fluctuating active power and compensating reactive power. The control results depend on the values of scaling factors and membership functions in fuzzy reasoning. Therefore, it is desired to obtain better control results that the scaling factors and membership functions are successively adjusted according to the load power fluctuation. In this paper, a new control strategy of leveling load power fluctuation by fuzzy-neural network with auto-acquiring of membership functions is proposed. Neural networks are used for an acquiring method of membership functions in fuzzy reasoning rules and auto-tuning of scaling factors and grade of fuzzy reasoning.