Cost-based Hypothetical Reasoning is an important framework for knowledge-based systems; however it is a form of non-monotonic reasoning and thus an NP-hard problem. To find a near-optimal solution in polynomial time with respect to problem size, some algorithms have been developed so far using optimization techniques. In this paper, we show two major ways of transforming propositional clauses in the hypothetical reasoning problems into constraints. One transforms the clauses into linear inequalities and is good at getting a low-cost solution, while the other transforms them into non-linear equalities and is good at finding a feasible solution. We them show a method of integrating these two transformations by using augmented Lagrangian method. Here, each variable and constraint is regarded as a processor and the searh is realized by their interaction. Two kinds of processors are derived from the two transformations; the structure of these processors are changed dynamically during the search. The cooperation of these two processors allows to obtain better near-optimal solutions than by our previous SL method. This effect is shown in the experiments using two problems with different problem structures.

抄録全体を表示

In order to utilize a large quantity of information in Internet, machine processing of HTML documents has been becoming tremendously important. HTML, however, is designed mainly for reading with browsers, thus not suitable for machine processing. XML was proposed as a solution for this problem. Unfortunately, full automatic transformation from HTML to XML is extremely difficult, because it absolutely demands to understand the meaning of HTML documents. On the other hand, there are many series of HTML pages in actual Web sites. Each page of a series usually has a quite similar structure with each other. Therefore a case-based transformation must be a promising method in practice. In this paper, we give a case-based transformation method from HTML documents to XML ones. Given a series of HTML documents and a sample transformation from a selected HTML document into XML one, we first analyze both of the semantic and syntactic information appearing in the sample pair. Next the remaining HTML pages of the series are automatically transformed into XML documents by using the information previously extracted from the sample. We adopt a vector model of term weighted frequency for approximating the meaning of HTML documents, and also use both headlines and a parse tree as syntactical information. Throughout experimental evaluation, we show this case-based method achieved a highly accurate transformation, i.e., 80% of actual 80 pages can be transformed in a correct way.

抄録全体を表示

AdaBoost has been successfully applied to a number of classification tasks, seemingly defying problems of overfitting. AdaBoost performs gradient descent in an error function with respect to the margin. This method concentrates on the patterns which are hardest to learn. However, this property of AdaBoost can be disadvantageous for noisy problems. Indeed, theoretical analysis has shown that the margin distribution plays a crucial role in understanding this phenomenon. Loosely speaking, some outliers should be tolerated if this has the benefit of substantially increasing the margin on the remaining points. In this paper, we propose new noise robust boosting methods using the concepts of ν-Support Vector Classification and Arc-GV. These methods allow for the probability of a pre-specified fraction ν of points to lie in the margin area or even on the wrong side of the decision boundary. This algorithms can give a nicely interpretable way of controlling the trade-off between minimizing the training error and capacity.

抄録全体を表示

In the facial retrieval, it is necessary to cope with the change of expression and consider the personal subjectivity. Generally, the matching by using linguistic keywords or image has been used in the conventional image retrieval. Therefore, this method isn’t suitable for the facial image retrieval. Recently, a method treating the user’s subjectivity is researched. In the method, the user evaluates facial images presented by the system, and the system retrieves new candidacies according to the evaluation. One of such method is the simulated breeding. However, this method gives heavier work to the user than the conventional method does. In this paper, we propose the facial image retrieval system united the retrieval by using linguistic keywords with the simulated breeding in order to decrease user’s work. The proposed system performs the retrieval by using the impressive adjective first, and the retrieval by using the simulated breeding. We evaluate the system through the experiment using the subjects and comparing with the prior art.

抄録全体を表示

Development of intelligent robots and rapid increase of aged societies have brought serious necessity of such systems that should facilitate mutual translation of sensory data and linguistic expressions. They are expected to help people, especially with some defected sense-organ, by translating sensory data into words such as “Pungent smell is sensed in the refrigerator!”, and otherwise enable people to order a robot to work by words such as “Search the room for a varicolored object.” For the purpose to develop such a system, the authors have recently analyzed and described the concepts of Japanese words for “color” and “lightness” based on MIDST(Mental image directed semantic theory) proposed by Yokota, M. et al. The analysis and description of the word concepts were performed approximately in the process as follows. Firstly, each word representing a specific color or lightness(“赤(=red)”, “暗(=dark)”, etc) was associated with a set of specific coordinates (point or range) of the color solid and its concept was defined as such a set of coordinates. Secondly, the words concerning temporal change or spatial distribution of color or lightness(“赤らむ(=redden)”, “ぼかす(=gradate)”, etc) were described as spatiotemporal relations among coordinates of the color solid. Thirdly, a computer system working with image input devices was constructed in order to ground words on real sensory data of color and lightness via the coordinates of the color solid in an interactive way with a human instructor, and has been found a fairly good success.

抄録全体を表示

This paper presents an approximate algorithm for the winner determination problem in combinatorial auctions, which is based on limited discrepancy search (LDS). Internet auctions have become an integral part of Electronic Commerce and can incorporate large-scale, complicated types of auctions including combinatorial auctions, where multiple items are sold simultaneously and bidders can express complementarity among these items. Although we can increase participants’ utilities by using combinatorial auctions, determining the optimal winners is a complicated constraint optimization problem that is shown to be NP-complete. We introduce the idea of LDS to an existing algorithm based on A^*. The merit of LDS is that it can avoid time-consuming re-computation of heuristic function h(·), since LDS is less sensitive to the quality of h(·). It can also limit the search efforts only to promising regions. Experiments using various problem settings show that, compared with the existing optimal algorithm, LDS can find near-optimal solutions (better than 95%) very quickly (around 1% of the running time).

抄録全体を表示