All issues
Volume 35, Issue 11
Displaying 1-23 of 23 articles from this issue
- |<
- <
- 1
- >
- >|
-
Theoretical Analysis of the Unimodal Normal Distribution Crossover for Real-coded Genetic AlgorithmsHajime KITA, Isao ONO, Shigenobu KOBAYASHI1999 Volume 35 Issue 11 Pages 1333-1339
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSReal-coded genetic algorithms attract attention as global optimization methods for nonlinear functions. For real-coded genetic algorithms, there have been proposed many crossover operators so far. Among them, the unimodal normal distribution crossover (UNDX) developed by One et al. shows good performance in optimization of multi-modal and highly epistatic fitness functions. However, the perfomance of the crossover operators have been evaluated only through numerical experiments with some benchmark problems, and clear guidelines to design operators have not been established.
In this paper, first, statistical characteristics of the UNDX are discussed theoretically. The results of the analysis show that the UNDX inherits the statistics of the parent population such as the mean vector and the variance-covariance matrix well. Based on this finding, the authors propose several guidelines to design crossover operators for the real-coded genetic algorithms.View full abstractDownload PDF (2881K) -
Its Verification with Cellular AutomataAtsunobu ICHIKAWA1999 Volume 35 Issue 11 Pages 1340-1345
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThe system generating element (SGE), that is the minimal functional element of emerging systems, is to have the following attributes:
-the SGE is of self-replicating,
-the SGE has chances to change its attributes,
-the existence of a SGE causes the change of the rate of replication of other SGEs.
Infinite state cellular automata composed of cells playing SGE are chosen as the model for verification of the role of the SGE. The numerical experiments with the model show that the systems really emerge out of a SGE, that is,
-the complex elements of system emerge and
-the interactions among the elements emerge.
The verification tells us that the existence of the SGE is also the condition of self-organization of systems.
Some of the characteristics of emerging process observed with the model are to be seen in real-life systems such as in the biological system. This observation suggests us that we need to differentiate the characteristics from specific ones adjunct to the specific real-life systems. In other words, they are not specific to the real-life system but rather universal to the emerging system.View full abstractDownload PDF (1490K) -
Learning of Dynamic Communication Using a Recurrent Neural NetworkKatsunari SHIBATA, Koji ITO1999 Volume 35 Issue 11 Pages 1346-1354
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSWe believe that communication in multi-agent system has two major meanings. One of them is to transmit one agent's observed information to the other. The other meaning is to transmit what an agent is intending. In such communication, dynamic communication with a communication loop is required. Here we focus the latter, and aim to the emergence of the autonomous and decentralized arbitration through communication among some agents. The communication contents and representation of them are not prescribed and are acquired by learning using a reinforcement signal, which is given to the agent after its action. The reinforcement signal is not shared with the other agents. Since the agent often has to make a decision from the past communication signals, the architecture using recurrent type (Elman) neural network is proposed. The ability of this architecture was examined by two and four agent negotiation problems. A variety of negotiation strategies (individuality) to avoid the conflicts after their decisions emerged among them through the simple learning with the simple architecture.View full abstractDownload PDF (6040K) -
An Application to the Graph Coloring ProblemKiyoharu TAGAWA, Kenji KANESIGE, Katsumi INOUE, Hiromasa HANEDA1999 Volume 35 Issue 11 Pages 1355-1362
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSIn this paper, a meta-heuristic method that combines the global search power of Genetic Algorithms with the local search power of local optimization algorithms is described. First of all, a metric function between two solutions, or phenotypes, is defined by the shortest Hamming distance between sets of isomorphic genotypes. The phenotypic distance is useful to analyze and control the behavior of genotypes in the search space from the view point of the problem space. Then, by using the phenotypic distance, a new crossover technique named Harmonic Crossover is proposed in which children always come to the position between their parents in the problem space. Finally, in order to get maximum efficiency of the meta-heuristic method, the best timing in employing the local optimization algorithms is discussed. The experimental results indicate that the local optimization algorithms should be synchronized with the Harmonic Crossover to find high quality solutions in reasonable time.View full abstractDownload PDF (3008K) -
Makoto OCHI, Manabu KOTANI, Kenzo AKAZAWA1999 Volume 35 Issue 11 Pages 1363-1369
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThere are some studies that genetic programmings have been used for creations of discriminant functions. However, these studies treated tree structure and it was very difficult to understand the evolved tree structure. Furthermore, the amounts of calculation increased very much. In this paper, we propose a method for evolution of the discriminant function using genetic algorithms and examine its performance for various experiments. We assume that the discriminant function can be approximated by polynomial expressions that are the product sum of input variables. Terms of the function are searched by genetic algorithms and the coefficients of each term are calculated by the multiple regression analysis. From experimental results, we comfirmed that the proposed method was effective for the creation of the discriminant function.View full abstractDownload PDF (3628K) -
Nobuo SANNOMIYA, Ya-jie TIAN, Hiroshi NAKAMINE1999 Volume 35 Issue 11 Pages 1370-1376
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSA fish school is considered as a typical example of autonomous decentralized system existing in nature, because it often shows a high degree of cooperation in the absence of a leader. In this paper simulation and analysis of the behavior of a fish school are made for the case where the school is affected by a trap. A specific order of the school is established based on the environmental effect and the information exchange among the school members. The relationship between the order of the school and the quantity of the information exchange is investigated by using two fish behavior models; a homogeneous cooperative school model and a heterogeneous repulsive school model. The different moving patterns are found as the simulation results by changing the quantity of information exchange for two types of fish bahavior models. It is also found that the emergence condition of cooperation is related to the system diversity.View full abstractDownload PDF (1559K) -
Daisuke KURABAYASHI, Jun OTA, Tamio ARAI, Katsuyuki NOGUCHI1999 Volume 35 Issue 11 Pages 1377-1384
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSIn this paper, we propose a heterogeneous robot group for cooperative task execution. The “heterogeneous” robot group is composed of autonomous robots which have different motion algorithms. In this paper, we investigate characteristics of the heterogeneous robot groups by simulations of multi-agent taveling salesman problem, which we can hardly solve by centralized algorithm. Then we propose an algorithm to arrange balance of mixture of robots driven by different algorithms. We verify the efficiency of the autonomous heterogeneous robot group by simulations.View full abstractDownload PDF (1731K) -
Masakazu SUZUKI, Masahiko FUCHI1999 Volume 35 Issue 11 Pages 1385-1393
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThe multi-stage genetic algorithm is proposed to solve a class of large scale optimization problems. The original problem with complicated constraints is divided into a supervisary problem and local sub-problems with simple constraints. Every sub-problem is solved by GA to generate a set of suboptimal solutions. And in the supervisary problem, the elements of each set are optimally combined by GA to yield the optimal solution for the original problem. The empirical knowledge obtained by solving the problem is effectively utilized to solve similar problems. The method is thus regarded as a learning solution. The extended knapsac problem is formulated and solved as an example to demonstrate the proposed method, and the efficiency of the method is shown. In addition, the method is successfully applied to optimal operation planning problem at LNG(Liquefied Natural Gas) terminal.View full abstractDownload PDF (1678K) -
Hirosuke HORII, Susumu KUNIFUJI, Teruo MATSUZAWA1999 Volume 35 Issue 11 Pages 1394-1399
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThere are two typical problems in Genetic Algorithms (GAs). First, GAs require huge calculation time for the genetic operations, such as selection, crossover, mutation, and individuals' fitness evaluations. Secondly, maintenance of the diversity of the population is necessary to avoid the premature convergence which spreads local optimum solution and stagnates the evolution. Island Parallel GAs divide a population into subpopulations and assign them to processing elements on a parallel computer in order to improve the processing speed. Each subpopulation searches for the optimal solution independently, and exchanges individuals periodically in order to maintain the diversity of each subpopulation. This exchange operation is called migration. In this research, we propose a new migration scheme. Individuals are exchanged among subpopulations asynchronously according to each subpopulation's search situation. The effect of the new migration scheme on the combinational optimization problems was verified by applying our algorithm to Knapsack Problems and Royal Road Functions using a parallel computer CRAY-T3E. Through these experiments, the following results were obtained. The migration scheme proposed in this research is effective for the combinational optimization problems. Especially, when our algorithm is applyed to the problem which has strong tendency of building block hypothesis, our algorithm performs effectively by the parallel search of the building blocks.View full abstractDownload PDF (1792K) -
Shinichiro YOSHII, Yukinori KAKAZU1999 Volume 35 Issue 11 Pages 1400-1406
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSComplex adaptive system is a comprehensive notion for living systems. The basis of this notion emphasizes interactions taking place between a function and information. In particular, to further our understanding of mechanism of complex adaptive systems, attention should be given to the way of modeling free from any preceding framework defined in advance by an arbitrary description which will result in restricting its capability to allow new interactions to emerge. Thus, this paper discusses several essential issues for modeling a universal complex adaptive system and proposes an ecological model based on the viewpoint of the internal measurement. The model consists of interacting schemes described in the form of tapes for a universal Turing machine. The internality perspective allows constituents to interact with each other highlighting their individuality. In that model, it is not axiomatic even whether an interaction will halt due to its characteristic arising from the computational universality. However, computer simulations show emergent dynamics where the semantics of the schemes are bifurcated into functions of measurement and descriptions of information through their continuous local interaction. Although the system is not provided with any particular mechanism for self-reproduction or evolution, the system is self-organized so that its schemes can utilize their environments for survival. Using this ecosystem model, this paper discusses how an emergent information process is achieved and a prospective for the realization of a system capable of emerging its functions depending on its environment.View full abstractDownload PDF (4818K) -
Proposal of a Neural Network with Dynamically-Rearranging FunctionToshiyuki KONDO, Akio ISHIGURO, Yoshiki UCHIKAWA, Peter EGGENBERGER1999 Volume 35 Issue 11 Pages 1407-1414
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSRecently, the Evolutionary Robotics approach has been attracting a lot of concern in the field of robotics and artificial life. In this approach, neural networks are widely used to construct controllers for autonomous mobile agents, since they intrinsically have generalization, noise-tolerant abilities and so on. However, the followings are still open questions: 1) the gap between simulated and real environments, 2) the evolutionary and learning phase are completely separated, and 3) the conflict between stability and evolvability/adaptability. In this paper, we try to overcome these problems by incorporating the concept of dynamic rearrangement function of biological neural networks with the use of neuromodulators. Simulation results show that the proposed approach is highly promising.View full abstractDownload PDF (4126K) -
Kotaro HIRASAWA, Junichiro MISAWA, Jinglu HU1999 Volume 35 Issue 11 Pages 1415-1420
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSRecently many researches on group robot systems have been studied, where a number of robots behave in a group like birds' or ants. It is generally known that each robot has a limited intellectual power, but the robots can behave more intellectually in a group because they can interact each other. One of the most famous researches in these fields is Boids which is the artificial model of the birds behavior in the computer software. And there have been reported the multi-agent robot systems which can do many kinds of tasks efficiently by training the rules between environments and actions using reinforced learning. This paper also proposes a multi-agent system where a criterion function is defined regarding the behavior of the multi-agent system and parameters of mutual interaction of the agents are trained in order to optimize the above criterion function. From simulations, it has been shown that emergent behaviors of the agents can be developed by appropriately adjusting the parameters.View full abstractDownload PDF (1990K) -
Eiichi YOSHIDA, Satoshi MURATA, Shigeru KOKAJI, Kohji TOMITA, Haruhisa ...1999 Volume 35 Issue 11 Pages 1421-1430
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSA distributed self-assembly and self-repair method is proposed for a 3-D reconfigurable machine, composed of many identical mechanical units. The method aims to enable the machine to transform itself into desired structure from an arbitrary initial configuration. The proposed method is implemented in such a way that each unit has identical software, so that any unit can play any role in the system. The method consists of two parts, for small and large scale systems. The first part is featured by a stochastic relaxation process, which allows the system to converge to a given target structure by searching for a proper unit motion over many degrees of freedom. We have also developed another self-assembly and self-repair method dedicated to effective self-assembly and self-repair for large scale systems by using layered and recursive description of target shape. The effectiveness of the proposed method is verified through a series of computer simulations.View full abstractDownload PDF (6120K) -
Ryota NOGUCHI, Motoaki MATSUZAKI, Ryotaro KOBAYASHI, Hideki ANDO, Tosh ...1999 Volume 35 Issue 11 Pages 1431-1437
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSA variety of branch predictors was previously proposed, with the analysis of branch behavior. To further improve the branch prediction accuracy, it is necessary to analyze branch behavior more deeply and feed it back to the function of branch predictors. This work, however, is very difficult. Thus, automating this work is helpful to reduce this difficulty. Although there exists an previous work for this automation using genetic algorithm, it was not successful. We attempted to generate a logic circuit of a practical branch predictor with high prediction accuracy by exploiting the existing framework of two-level branch predictors. Our results show that the branch predictor generated by genetic algorithm can achieve 0.5% higher prediction accuracy than the gshare.View full abstractDownload PDF (1792K) -
Hisashi HANDA, Osamu KATAI, Norio BABA, Tetsuo SAWARAGI, Tadataka KONI ...1999 Volume 35 Issue 11 Pages 1438-1446
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSIn this paper, we propose a new Genetic Algorithm involving a mechanism of Co-evolution. Our Coevolutionary Genetic Algorithm consists of two Genetic Algorithms (GAs): a traditional GA which searches for the optimal solutions for given problems, and another GA to search for effective schemata in the former GA. Also, we adopt binary coding, partial spaces in genetic space, and rotated partial spaces as the coding method for the individuals in the latter GA, namely, schemata for the former GA. Moreover, we discuss on the effectiveness of coding and fitness evaluation for our Coevolutionary Genetic Algorithm. Several computational results on function optimization problems confirm the effectiveness of our approach.View full abstractDownload PDF (4610K) -
Hideo YUASA, Masami ITO1999 Volume 35 Issue 11 Pages 1447-1453
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSSome kinds of spatiotemporal pattern generator systems are expressed by evolution equations. Such evolution equations are composed of many dynamic units which behave from only its local information. This is one example to coordinate many subsystems which observe the system from inside and to generate a global order. This paper shows how to treat these evolution equations and how to apply it to network systems. First, a continuous media system which is expressed by a gradient system in function space is considered. One of the simplest potential functional derives a reaction diffusion equation which decreases the value of potential functional monotonously. The similar result is found in graph space. That is, a reaction diffusion equation on a graph decreases the value of a potential functional monotonously. This means that a network of dynamic units which observe only their connecting units' states can generate a global order in the same way of continuous media. This theory can treat some internal dynamic network system which should coordinate without some kinds of central controllers.View full abstractDownload PDF (2215K) -
Setsuya KURAHASHI, Ushio MINAMI, Takao TERANO1999 Volume 35 Issue 11 Pages 1454-1461
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThis paper proposes a new method: Inverse Simulation for analyzing emergent behaviors of agents in artificial societies, which model social interactions in the electronic mediated communication. Inverse Simulation utilizes Genetic Algorithms with tabu search to optimize a global evaluation function. The method is implemented in a simulator TRURL, which evolves artificial worlds of multi-agents to socially interact with each other. The micro-level agent activities are determined by both predetermined and acquired parameters. The former pa-rameters have constant values during one simulation cycle, however, the latter parameters change during the interactions. Unlike conventional artificial society models, TRURL evolves the societies by changing the predetermined parameters to optimize macro-level socio-metric measures, which can be observed in such real societies as e-mail oriented organizations and electronic commerce markets. Thus, using TRURL, we automatically tune the parameters up and observe both micro- and macro-level phenomena grounded in the activities of real worlds.View full abstractDownload PDF (3288K) -
Teijiro ISOKAWA, Nobuyuki MATSUI, Haruhiko NISHIMURA1999 Volume 35 Issue 11 Pages 1462-1468
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSGenetic Algorithm is well-known as the optimizing algorithm taken after evolutionary strategy, and widely used for engineering problems that occur when designing a system by its self-organizing phenomenon. It is based on Neo-Darwinism, that is, the theory that individuals with advantageous characters to the given environment increase and individuals with disadvantageous ones decrease when mutation occurs in genetic characters in a population. In Genetic Algorithm, there is a premise that all the genetic characters in an organism have some meanings of adaptation. In the natural organisms, however, there are a number of neutral mutations, which is neither advantageous nor disadvantageous to their adaptation. This fact suggests that there is a new possibility of genetic mechanism in designing evolutionary systems. This is the framework motivated by the neutral theory of molecular evolution, which is different from the conventional design based on Neo-Darwinian Genetic Algorithm. In the neutral theory, some genetic characters in an individual are not positively possessed under selection pressure (not by orthogenesis), but allowed to be fixed by chance, remaining neutral to selection pressure (by random genetic drift). This leads to the supplement of redundancy to genetic information which represents an individual, and the appearance of the diversity of genetic information in a population. In this paper, in order to investigate the above point concretely, we adopt the Ladder-Network as a simple model which makes permutation of information. When we evaluate the fitness by using only the degree of correspondence between target alignment and output one, the factor of the number of steps in the network becomes the neutral character which does not directly affect the fitness.View full abstractDownload PDF (3815K) -
Kazuhiro OHKURA, Yoshiyuki MATSUMURA, Kanji UEDA1999 Volume 35 Issue 11 Pages 1469-1477
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSEvolution strategies (ESs) constitute a class of engineering optimization algorithms based on the model of natural evolution. ESs are used particularly for real-valued function optimization in which they show better performance than the other evolutionary algorithms in many problems. However, as shown in this paper, their performance dramatically changes according to the lower bound of strategy parameters, although they are traditionally considered to be controlled by a so-called “self-adaptive” property of their own. Therefore, ESs should be applied to each optimization problem with a carefully selected lower bound. In order to overcome this brittleness, this paper proposes a new extended ES called Robust ES (RES). RES adopts redundant individual representation and new mutation mechanisms so that the strategy parameters can be changed by not only their self-adaptive mechanisms based on natural selection but also the effect of genetic drift in their non-coding region. Computer simulations using several test functions are conducted to illustrate the robustness of the proposed approach.View full abstractDownload PDF (1974K) -
Kazuhiro KOJIMA, Koji ITO1999 Volume 35 Issue 11 Pages 1478-1485
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSWe proposes an autonomous dynamical pattern recognition and learning system. It is demonstrated that, first, when the embedded pattern, i.e., known pattern, is given to the network, the firing pattern of the network immediately goes to the relevant embedded pattern and the network state reduces to the oscillatory state at once. Second, when no embedded pattern, i.e., unknown pattern, is given to the network, the network state oscillates chaotically. It is considered as “I don't know” state. Finally, when Hebb rule is applied to the network under the external stimuli that are unknown patterns, the internal state of the network is inversely bifurcated from the chaotic state to the periodic state according to the progress of learning. By utilizing this phase transition as an index of the progress of learning, the network can learn new patterns without any external observers.View full abstractDownload PDF (4245K) -
Keiki TAKADAMA, Takao TERANO, Katsunori SHIMOHARA, Koichi HORI, Shinic ...1999 Volume 35 Issue 11 Pages 1486-1495
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSThis paper investigates the effectiveness of an emergent problem solving method which introduces the characteristics of multiagent learning analyzed from the viewpoint of both organizational learning (OL) in social science and genetics-based machine learning (GBML). A careful investigation of the above method has revealed the following implications: (1) there are two levels in the learning mechanisms of multiagent learning (the indivictual and organizational level) and each level is divided into two types (single- and double-loop learning). The integration of these four learning mechanisms improves the collective performance (good solution with less computational cost) in multiagent environments; (2) the effectiveness of the emergent problem solving in multiagent environments is supported by the following three properties: (a) different dimension in learning mechanisms, (b) meta-level interaction in addition to the interaction among agents, and (c) a combination of exploration at an individual level andt exploitation at an organizational level.View full abstractDownload PDF (2611K) -
Osamu KITAURA, Hideaki ASADA, Motoaki MATSUZAKI, Takamitsu KAWAI, Hide ...1999 Volume 35 Issue 11 Pages 1496-1504
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSGenetic algorithms (GAs) are effective for large-scale optimization problems. Several GA engines that reduce computation time have been proposed. Although these engines accelerate execution of GAs over software implementations, the speedup is not enough. This problem arises from less considerations to an efficient pipeline design. The pipeline stalls over the most of the execution time. We propose a new architecture of a GA engine, which we call H3 engine, whose pipeline never stalls. To remove all of the pipeline stalls, our H3 engine employs steady stat e GA and pipelines the roulette wheel selection using the combination of binary search and linear search. We implement H3 engine on an FPGA and evaluate its performance. Our evaluation results show that H3 performs GAs about 730 times faster than software. We also discuss implementation of H3 engine for large-scale application s.View full abstractDownload PDF (2890K) -
Tetsuo SAWARAGI, Naoki TANI1999 Volume 35 Issue 11 Pages 1505-1513
Published: November 30, 1999
Released on J-STAGE: March 27, 2009
JOURNAL FREE ACCESSA personal learning apprentice system such as an interface agent has to be able to selectively pick up regularities contained in a stream of actual observations as well as be able to construct a user's concept by actively inferring what the user is supposed to perceive from his/her apparent behaviors. We propose a conceptual learning method for such an agent using an evolutinary computation. Our proposed algorithm comprises two processes; adaptive feature selection and GA-based feature discovery. The former selects the essential attributes out of a provided set of attributes that may initially be either relevant or irrelevant, and the latter constructs new attributes using genetic algorithms applied to a set of elementary features logically represented in a disjunctive normal form. Our method cab be applied to artificial data as well as to a data set obtained from human-machine interactions observed during operation of a simulator of a generic dynamic production process.View full abstractDownload PDF (4477K)
- |<
- <
- 1
- >
- >|