Bootstrapping has a tendency, called semantic drift, to select instances unrelated to the seed instances as the iteration proceeds. We demonstrate the semantic drift of Espresso-style bootstrapping has the same root as the topic drift of Kleinberg's HITS, using a simplified graph-based reformulation of bootstrapping. We confirm that two graph-based algorithms, the von Neumann kernels and the regularized Laplacian, can reduce the effect of semantic drift in the task of word sense disambiguation (WSD) on Senseval-3 English Lexical Sample Task. Proposed algorithms achieve superior performance to Espresso and previous graph-based WSD methods, even though the proposed algorithms have less parameters and are easy to calibrate.

抄録全体を表示

In this paper, we propose a video scene annotation method based on tag clouds. First, user comments associated with a video are collected from existing video sharing services. Next, a tag cloud is generated from these user comments. The tag cloud is displayed on the video window of the Web browser. When users click on a tag included in the tag cloud while watching the video, the tag gets associated with the time point of the video. Users can share the information on the tags that have already been clicked. We confirmed that the coverage of annotations generated by this method is higher than that of the existing methods, and users are motivated to add tags by sharing tag clouds. This method will contribute to advanced video applications.

抄録全体を表示

The two subtasks of predicate-argument structure analysis -- argument role classification and predicate word sense disambiguation, are mutually related. Information of argument roles is useful for predicate word sense disambiguation, at the same time, the predicate sense information can be an important clue for argument role labeling. However, most of the existing approaches do not model such structural interdependencies. In this paper, we propose a structured prediction model that learns predicate word senses and argument roles simultaneously. In order to deal with the structural interdependencies, we introduce two factors: pairwise factor that captures local dependencies between predicates and arguments, and global factor that captures non-local dependencies over whole predicate-argument structure. We propose a new large-margin learning algorithm for linear models, in which the global factor is handled in parallel with the local factor. In the experiments, the proposed model achieved performance improvements in both tasks, and competitive results compare to the state-of-the-art systems.

抄録全体を表示

Automatic construction methods for image processing proposed till date approximate adequate image transformation from original images to their target images using a combination of several known image processing filters by evolutionary computation techniques. Genetic Image Network (GIN) is a recent automatic construction method for image transformation. The representation of GIN is a network structure. In this paper, we propose a method of automatic construction of image classifiers based on GIN, designated as Genetic Image Network for Image Classification (GIN-IC). The representation of GIN-IC is a feed-forward network structure. GIN-IC is composed of image transformation nodes, feature extraction nodes, and arithmetic operation nodes. GIN-IC transforms original images to easier-to-classify images using image transformation nodes, and selects adequate image features using feature extraction nodes. We apply GIN-IC to test problems involving multi-class categorization of texture images and two-class categorization of pedestrian and non-pedestrian images. Experimental results show that the use of image transformation nodes is effective for image classification problems.

抄録全体を表示

In computer vision, many systems for video surveillance have been studied and developed. Finding appropriate features is very important especially for robust object tracking. This paper proposes a new system employing an evolutionary technique to track moving objects such as pedestrian effectively. In the proposed system, a number of agents are generated for each object, and independently search and move to the predicted position from features of local region. Each agent can select the appropriate feature relevant for the scene and the object to track by an evolutionary method using two values of feature effectiveness: normality and separateness. We carried out experiments using some scenes in an outside parking. The proposed system showed much better performance compared with the conventional system and a system using particle filter.

抄録全体を表示

In this paper, we focus on analyzing the behavior of the selection models for real-coded genetic algorithms. Recent studies show that Just Generation Gap (JGG) selection model outperforms Minimal Generation Gap (MGG) model when a multi-parental crossover operator based on the hypothesis of the preservation of the statistics of parents is used. However, the validation of JGG selection model is not done yet. To validate the selection method of JGG, we analyze the differences of the behavior of JGG selection model and that of MGG selection model.

抄録全体を表示

The purpose of this paper is to propose a new real-coded genetic algorithm (RCGA) named Networked Genetic Algorithm (NGA) that intends to find multiple optima simultaneously in deceptive globally multimodal landscapes. Most current techniques such as niching for finding multiple optima take into account big valley landscapes or non-deceptive globally multimodal landscapes but not deceptive ones called UV-landscapes. Adaptive Neighboring Search (ANS) is a promising approach for finding multiple optima in UV-landscapes. ANS utilizes a restricted mating scheme with a crossover-like mutation in order to find optima in deceptive globally multimodal landscapes. However, ANS has a fundamental problem that it does not find all the optima simultaneously in many cases. NGA overcomes the problem by an adaptive parent-selection scheme and an improved crossover-like mutation. We show the effectiveness of NGA over ANS in terms of the number of detected optima in a single run on Fletcher and Powell functions as benchmark problems that are known to have multiple optima, ill-scaledness, and UV-landscapes.

抄録全体を表示

Applications of memetic algorithms (MAs) are usually computationally expensive. In this paper we suggest efficient search limiting strategies for local search used in MAs because local search is the most time consuming part of MAs. The suggested strategies are applied to a recently proposed powerful MA for the capacitated vehicle routing problem (CVRP). Experimental results on the well-known benchmarks show a significant speed-up of 80% in running time without worsening the solution quality. Moreover, the MA dominates state-of-the-art heuristics for the CVRP with respect to both the computation time and the solution quality.

抄録全体を表示

In this paper, we propose a new method to obtain the transition rules of two-dimensional cellular automata (CA) that performs grayscale image processing. CA has the advantages of producing complex systems from the local interaction of simple elements, and has attracted increased research interest. The difficulty of designing CA's transition rules to perform a particular task has severely limited their applications. So, the evolutionary design of CA rules has been studied. In this method, an evolutionary algorithm was used to evolve CA. In recent years, this method has been applied to image processing. Rosin has studied the evolutionary design of two-dimensional CA to perform noise reduction, thinning and convex hulls. Batouche et al. and Slatnia et al. employed genetic algorithm to investigate the possibility of CA to perform edge detection. In the previous methods, 2-state CA was used for binary image processing. Unlike the previous methods, the present method uses 256-state CA rules to perform grayscale image processing. Gene Expression Programming (GEP) proposed by Ferreira is employed as a learning algorithm in which the chromosomes encode the transition rules as expression trees. Experimental results for the reduction of impulse noise, salt-and-pepper noise and gaussian noise show that the proposed method is equivalent to previous methods in performance and more than 100 times faster than the method proposed by Rosin. We show that the rule obtained by the proposed method employs symmetry-based strategy in the noise reduction process and this property can reduce complexity of CA.

抄録全体を表示

In this work, we propose a novel multi-objective evolutionary algorithm (MOEA) which improves search performance of MOEA especially for many-objective combinatorial optimization problems. Pareto dominance based MOEAs such as NSGA-II and SPEA2 meet difficulty to rank solutions in the population noticeably deteriorating search performance as we increase the number of objectives. In the proposed method, we rank solutions by calculating Pareto partial dominance between solutions using r objective functions selected from m objective functions to induce appropriate selection pressure in many-objective optimization by Pareto-based MOEA. Also, we temporally switch r objective functions among _mC_r combinations in every interval generations I_g to optimize all of the objective functions throughout the entire evolution process. In this work, we use many-objective 0/1 knapsack problems to show the search performance of the proposed method and analyze its evolution behavior. Simulation results show that there is an optimum value for the number of objective functions r to be considered for the calculation of Pareto partial dominance and the interval (generation numbers) I_g to maximize the entire search performance. Also, the search performance of the proposed method is superior to recent state-of-the-art MOEAs, i.e., IBEA, CDAS and MSOPS. Furthermore, we show that the computational time of the proposed method is much less than IBEA, CDAS and MSOPS, and comparative or sometimes less than NSGA-II.

抄録全体を表示

Fitness landscapes which include neutrality have been conceptualized as containing neutral networks. Since the introduction of this concept, EC researchers have expected that a population can move along neutral networks without getting trapped on local optima. On the other hand, it has been demonstrated in tunably neutral NK landscapes that neutrality does not affect the ruggedness, although it does reduce the number of local optima. These show that the effects of neutrality are still contentious issues. This paper investigates the effects of neutrality and ruggedness on structures of fitness landscapes. A neutral network is described in a mathematical form based on Harvey's original definition with minor modifications. According to this description, genotypic search space in a problem which includes both neutrality and ruggedness can be classified into several neutral networks. These structures in and between neutral networks are then analyzed. Our results demonstrate that landscapes with a higher degree of neutrality have the larger sizes of neutral networks. For landscapes with the smallest degree of ruggedness, all neutral networks have some contact points to the networks of higher fitness. For landscapes with a higher degree of ruggedness, there are few contact points between the networks of high fitness and the ones of the highest fitness, which seem to be isolated, deceptive or rugged.

抄録全体を表示

In recent years, program evolution algorithms based on the estimation of distribution algorithm (EDA) have been proposed to improve search ability of genetic programming (GP) and to overcome GP-hard problems. One such method is the probabilistic prototype tree (PPT) based algorithm. The PPT based method explores the optimal tree structure by using the full tree whose number of child nodes is maximum among possible trees. This algorithm, however, suffers from problems arising from function nodes having different number of child nodes. These function nodes cause intron nodes, which do not affect the fitness function. Moreover, the function nodes having many child nodes increase the search space and the number of samples necessary for properly constructing the probabilistic model. In order to solve this problem, we propose binary encoding for PPT. In this article, we convert each function node to a subtree of binary nodes where the converted tree is correct in grammar. Our method reduces ineffectual search space, and the binary encoded tree is able to express the same tree structures as the original method. The effectiveness of the proposed method is demonstrated through the use of two computational experiments.

抄録全体を表示

As an approach for dynamic control problems and decision making problems, usually formulated as Markov Decision Processes (MDPs), we focus direct policy search (DPS), where a policy is represented by a model with parameters, and the parameters are optimized so as to maximize the evaluation function by applying the parameterized policy to the problem. In this paper, a novel framework for DPS, an exemplar-based policy optimization using genetic algorithm (EBP-GA) is presented and analyzed. In this approach, the policy is composed of a set of virtual exemplars and a case-based action selector, and the set of exemplars are selected and evolved by a genetic algorithm. Here, an exemplar is a real or virtual, free-styled and suggestive information such as ``take the action A at the state S'' or ``the state S1 is better to attain than S2''. One advantage of EBP-GA is the generalization and localization ability for policy expression, based on case-based reasoning methods. Another advantage is that both the introduction of prior knowledge and the extraction of knowledge after optimization are relatively straightforward. These advantages are confirmed through the proposal of two new policy expressions, experiments on two different problems and their analysis.

抄録全体を表示

This work proposes space partitioning, a new approach to evolutionary many-objective optimization. The proposed approach instantaneously partitions the objective space into subspaces and concurrently searches in each subspace. A partition strategy is used to define a schedule of subspace sampling, so that different subspaces can be emphasized at different generations. Space partitioning is implemented with adaptive epsilon-ranking, a procedure that re-ranks solutions in each subspace giving selective advantage to a subset of well distributed solutions chosen from the set of solutions initially assigned rank-1 in the high dimensional objective space. Adaptation works to keep the actual number of rank-1 solutions in each subspace close to a desired number. The effects on performance of space partitioning are verified on MNK-Landscapes. Also, a comparison with two substitute distance assignment methods recently proposed for many-objective optimization is included.

抄録全体を表示