This paper shows statistical analyses of the search-space landscape of travelling salesman problems in due consideration of stochastic optimization. It is known from existing works that travelling salesman problems have landscape called “a rugged landscape” and “big valley structure”. This work reveals more detailed structure of the landscape. We deal with the 1000 travelling salesman problems of 6 to 9 cities where the cities are arranged randomly and a travelling salesman problem of 100 cities. It is assumed that the rugged landscape is a combination of the global valleylike structure and the local noiselike structure. Each of them is characterized by the statistical properties of the search-space landscape, that is, the global valleylike structure has linearity with the distance (in this case, the bond distance) from the optimum, and the variance of the local noiselike structure increase monotonously with the distance from the optimum. On the other side correlation of the tours with the costs close upon the optimum cost is low. For this reason to combine the genetic search with the local search is supported. Even if the number of cities and the definition of the intercity cost value are changed, the structure of the landscape has the same feature. Although the number of the cities of the examined travelling salesman problems is not large, obtained results seem to be universal. It is forecasted that not only travelling salesman problems but also many practical problems have the structure which is characterized with the same measure. These results are useful to compose more effective optimization methods without trial and error.

抄録全体を表示

This paper proposes a method for discovering Web communities. A complete bipartite graph K_{i, j} of Web pages can be regarded as a community sharing a common interest. Discovery of such community is expected to assist users’ information retrieval from the Web. The method proposed in this paper is based on the assumption that hyperlinks to related Web pages often co-occur. Relations of Web pages are detected by the co-occurrence of hyperlinks on the pages which are acquired from a search engine by backlink search. In order to find a new member of a Web community, all the hyperlinks contained in the acquired pages are extracted. A page which is pointed by the most frequent hyperlinks is regarded as a new member of the community. We have build a system which discovers complete bipartite graphs based on the method. Only from a few URLs of initial community members, the system succeeds in discovering several genres of Web communities without analyzing the contents of Web pages.

抄録全体を表示

Crew Pairing is one of the most important and difficult problems for airline companies. Nets to fuel costs, the crew costs constitute the largest cost of airlines, and the crew costs depend on the quality of the solution to the pairing problem. Conventional systems have been used to solve a daily model, which handles only regular flights with many simplifications, so a lot of corrections are needed to get a feasible solution and the quality of the solution is not so high. A fully dated model, which handles regular flights and irregular flights simultaneously, is extremely hard and has not been solved directly hitherto. The number of irregular flights tend to increase in Europe and Japan, hence the resolution of the practical fully dated pairing system is desired. This paper presents a new approach which solves directly the fully dated crew pairing by Genetic Algorithms. GA is improved with stochastic processes to attack the fully dated model, and many heuristics are included in the decoding mecanism of the GA. For several hundred flights per day for fully dated 2 months scheduling period, the system found better or equivalent solutions to the human scheduler’s without any simplification from 5 to 20 times faster.

抄録全体を表示

This paper presents a new method that improves robustness of real-coded Genetic Algorithm (GA) for function optimization. It is reported that most of crossover operators for real-coded GA have sampling bias, which prevents to find the optimum when it is near the boundary of search space. They like to search the center of search space much more than the other. Therefore, they will not work on functions that have their optima near the boundary of the search space. Although several methods have been proposed to relax this sampling bias, they could not cancel whole bias. In this paper, we propose a new method, Toroidal Search Space Conversion (TSC), to remove this sampling bias. TSC converts bounded search space into toroidal one without any parameter. Experimental results show that a GA with TSC has higher performance to find the optimum near the boundary of search space and the GA has more robustness concerning the relative position of the optimum.

抄録全体を表示

This paper presents conceptual clustering algorithms using regression analysis. The basic idea is that given data can be classified to the class “existing” and so conceptual clustering(unsupervised learning) is transformed to classification (supervised learning). The algorithms consist of transforming given data to the data with a class, obtaining a function({0, 1}ⁿ →[0, 1]) by regression analysis, approximating the function by a Boolean function, and generating a concept hierarchy from the Boolean function. Regression analysis includes linear regression analysis and nonlinear regression analysis by neural networks. The algorithms can perform the multiple classification and generate simple clusters. The algorithms using linear regression analysis and neural networks have been applied to real data. Results show that the algorithm using neural networks works well.

抄録全体を表示

Estimating probability distributions on returns provides various sophisticated decision making schemes for control problems in Markov environments, including risk-sensitive control, efficient exploration of environments and so on. Many reinforcement learning algorithms, however, have simply relied on the expected return. This paper provides a scheme of decision making using mean and variance of returndistributions. This paper presents a TD algorithm for estimating the variance of return in MDP(Markov decision processes) environments and a gradient-based reinforcement learning algorithm on the variance penalized criterion, which is a typical criterion in risk-avoiding control. Empirical results demonstrate behaviors of the algorithms and validates of the criterion for risk-avoiding sequential decision tasks.

抄録全体を表示