It is not easy for a user of an Information Retrieval (IR) system to select an appropriate keyword set to represent his or her specific information need. Therefore, many IR systems can modify keyword sets by estimating the user's particular requirement. Even though such IR systems have better retrieval performance, the complicated estimation process entailed by a large number of keywords makes it difficult for a user to understand how the system behaves. Therefore, we used a thesaurus for query expansion. To select an appropriate keyword set, we proposed two concepts: ``adaptive generalization'', which estimates an appropriate generalization level of the given keywords by using relevant document information, and ``purpose-oriented concept structure modification'', which selects relevant keywords from a predefined synonym set in a thesaurus. Because query expansion based on a thesaurus aims to find new keywords that are complementary to the initial keywords, we proposed to use this method to construct a Boolean query formula to represent the user's information need. We proposed a new IR system called ``appropriate Boolean query reformulation for IR with adaptive generalization'' (ABRIR-AG) to support Boolean query formation. In ABRIR-AG, we reformulated a user-given Boolean query by using a small number of relevant documents. Finally, to evaluate its effectiveness, we evaluated ABRIR-AG by using a large-scale test collection containing WWW documents.
We propose a new information sharing system, named ``BisNet'', which automatically gathers information about the bookmarks stored in users' web browsers and helps the users exchange URIs of possibly interesting web pages with others who have similar interest with them. Being different from other typical agent services that gather and provide information according to pre-registered user profiles, BisNet is expected to share more relevant information because of its use of web browser bookmarks that are actively selected and ordered by many humans. To enhance the relevance of information being shared, we developed a novel algorithm for directory evaluation. This algorithm only looks at the local referential structure between bookmark directories and URIs and calculates for each directory the ``order index'' that represents how well its content URIs are put in order with a focus on specific areas of interest. Then each directory receives new URIs from other related directories with large order indexes. The repetition of such URI exchanges makes the whole directory-URI networks dynamically form directory groups according to the commonness of the URIs they refer to. Our method is unique in that it pays no attention to the actual contents of web pages, and thus is much simpler and faster than other methods based on the result of content analysis. We carried out a field trial that involved 45 people who used a prototype version of BisNet clients. The result indicated that the relevance of shared URIs positively correlated with the ``order index'' of surrounding related directories, demonstrating the effectiveness of the method we proposed.
We propose a parallel distributed interactive genetic algorithm(PDIGA) as a new design collaboration method. PDIGA uses an IGA and a parallel distributed genetic algorithm, which combines several IGA systems, and it has a scheme that best ones among the design solutions based on the subjective evaluation of each user are shared among users for every generation. A collaboration system using PDIGA is developed to make good design solutions among several people at difference locations. To verify a validity of the PDIGA system, we conducted experiments for comparing IGA and PDIGA. In particular, we examined each user's evaluation to the final design, compared the averages of the individual evaluation value, and examined the similarity of the design solution. These showed that the collaboration system using PDIGA become a consensus building system, and users' design solutions are unified in a group.
In the existing Reinforcement Learning, it is difficult and time consuming to find appropriate the meta-parameters such as learning rate, eligibility traces and temperature for exploration, in particular on a complicated and large-scale problem, the delayed reward often occurs and causes a difficulty in solving the problem. In this paper, we propose a novel method introducing a temperature distribution for reinforcement learning. In addition to the acquirement of policy based on profit sharing, the temperature is given to each state and is trained by hill-climbing method using likelihood function based on success and failure of the task. The proposed method can reduce the parameter setting according to the given problems. We showed the performance on the grid world problem and the control of Acrobot.
In this paper, we propose a novel usage for computational cognitive models. In cognitive science, computational models have played a critical role of theories for human cognitions. Many computational models have simulated results of controlled psychological experiments successfully. However, there have been only a few attempts to apply the models to complex realistic phenomena. We call such a situation ``open-ended situation''. In this study, MAC/FAC (``many are called, but few are chosen''), proposed by [Forbus 95], that models two stages of analogical reasoning was applied to our open-ended psychological experiment. In our experiment, subjects were presented a cue story, and retrieved cases that had been learned in their everyday life. Following this, they rated inferential soundness (goodness as analogy) of each retrieved case. For each retrieved case, we computed two kinds of similarity scores (content vectors/structural evaluation scores) using the algorithms of the MAC/FAC. As a result, the computed content vectors explained the overall retrieval of cases well, whereas the structural evaluation scores had a strong relation to the rated scores. These results support the MAC/FAC's theoretical assumption - different similarities are involved on the two stages of analogical reasoning. Our study is an attempt to use a computational model as an analysis device for open-ended human cognitions.