The capability of learning is one of the salient features of realtime search algorithms such as LRTA. These algorithms repeatedly perform problem solving trials so that the heuristic values will eventually converge to exact values along every optimal path to the goal. The major impediment is, however, the instability of the solution quality (the length of the solution path) during convergence. This instability is due to two properties of the search algorithms:(1) they try to find all optimal solutions even after obtaining fairly good solutions, and (2) they tend to move towards unexplored areas thus failing to balance exploration and exploitation. In this paper, we propose and analyze two new realtime search algorithms to stabilize the convergence process. ・ε-search (weighted realtime search) relaxes the condition of searching for optimal solutions to allow suboptimal solutions with ε error. As a result, ε-search significantly reduces the total amount of learning performed. ・ε-search (realtime search with upper bounds) utilizes the upper bounds of estimated costs, which become available after the problem is solved once. Guided by the upper bounds, δ-search can better control the tradeoff between exploration and exploitation. The ε-and δ-search algorithms can be combined easily. The effectiveness of these algorithms is demonstrated by solving randomly generated mazes.

The significance of functional representation has been pointed out by a lot of researchers, while the general relations between function and behavior is not fully understood yet. We consider a functional model of each component in a system as consisting of two elements. One is a necessary and sufficient information for simulating behavior of the component which we call a behavior model. The other is the interpretation result of the behavior under a desirable state which the component is expected to achieve, which we call function. By identifying primitives necessary for the interpretation of the behavior in various domains, we can capture what function is and represent it by selection and combination of them. As the representation of function and behavior plays an important role in modeling domain knowledge, identification of each of them contributes to organizing domain ontology. This paper proposes FBRL, a language for representing function and behavior with the primitives we identified and discuss its application to explanation generation. First half of this paper briefly explains FBRL. We also investigate the relation between function and behavior based on the primitives of FBRL. As FBRL can represent concepts at various levels of abstraction, it contributes to explanation generation by providing informations for mapping behavior of a component to a term which represents its function. Second half of this paper shows an example which shows how FBRL contributes to explanation generation. We conclude this paper with our next step toward the redesign task.

In this paper, we present a method and its implementation to aid the process of building knowledge out of chunks of information. The chunks of information, which can be the source of new knowledge, often exist in so called tacit dimension. This paper deals with these unclassified chunks of information that haven't been systematized yet but has potential to be systematized. In order to aid the process of building knowledge out of those chunks, we give a system which maps the chunks into a metric space. Through the reconfiguration of this space, the user can gradually structurize the chunks of information into systematized knowledge. The unique feature of the proposed system compared with the existing ones lies in the function to deal with the users point of view. The system reveals the user's hidden point of view explicitly, and this helps the process of systematization of knowledge much. We have implemented a system CASSIS, which stands for Computer-ASSIsted Systematization, and have carried out several experiments. We show the configuration of the system and the results of the experiments.

We have developed a deductive database system PACADE for analyzing three dimensional and secondary structure of protein. PACADE consists of an inference system and a relational database which stores data about primary, secondary and tertiary structure converted from Brookhaven's Protein Data Bank (PDB). In addition to usual searches for substructures of proteins, PACADE provides a similarity search function, based on similarity among the substructures. When a specific substructure is given as a closed query, the system returns a set of substructures similar to it. Through this operation, direct similarity relationships are obtained. Therefore, a closure of indirect similarity relationships can be computed by repeating this operation. The resulting closure is a cluster of substructures which includes the substructure given in the query, and there are indirect similarity relationships among them. Comparison of the clustering mentioned above with the clustering of proteins, based on biological functions, may lead to the discovery of structures concerning the biological functions of proteins. We show here that, by iterating a similarity search with a closed query, PACADE can automatically compute a closure of indirectly similar structures including the one in a query. The algorithm of the iterative similarity search provides a quick way to compute a closure of indirectly similar structures of proteins because: 1) It is not a combinatorial computation. 2) It performs differential evaluation analogous to the semi-naive evaluation in a bottom-up evaluator of deductive database system. 3) It employs Magic set transformation, which utilizes constant bindings in a closed query. Finally, we report results of experiments by computing a closure using secondary structure data stored in PACADE. The elapsed time for computation of a closure increased linearly to the size of search space, and it is a desirable feature under the situation that PDB data are exponentially increasing.

It is difficult to extract contents of images from image data itself. To avoid this difficulty, we intend to extract the content information of images from the explanation texts of image data. We select a pictorial book of flora as the explanation texts of image data. In that text, the properties of elements in image data (plants and its part) are often expressed in copular sentences. Therefore, we intend to extract the information about the properties of plants and its parts from copular sentences. To extract these information from the pictorial book of flora, we need technical knowledge of botany, namely, the relationship among technical terms in botany and the properties of plants and its parts. This technical knowledge is usually obtained by hand because it is difficult to obtain it automatically. In this paper, we described a new method of technical knowledge acquisition for information extraction from the pictorial book of flora. Our method is as follows. We extracted these copular sentences using simple pattern matching method, established relationship among the subject words of copular sentences manually, and classified the predicates of copular sentences automatically using a character-based best matching retrieval method and a method of finding coordinate structures of a sentence. We used this technical knowledge for the analysis of modifier phrases and the semantic analysis of copular sentences in the pictorial book of flora. We obtained the correct recognition scores of 96% in the analysis of modifier phrases and 87% in the semantic analysis of copular sentences.

Real-Time A^*(RTA^*) is an on-line search scheme in which look-ahead searches and moves are interleaved. It does not guarantee to find an optimal solution but it finds a semi-optimal solution with less search effort. To improve the quality of solution, Knight has proposed Multi-Agent Real Time A^*(MARTA^*), in which multiple agents concurrently and autonomously execute RTA^* for a given problem. In this scheme, agents do not coordinate their moves, but each agent just randomly chooses its move when the candidates look equally good. In this paper, we have interest in how agents should coordinate their moves to find a better solution faster. We propose two organizational approaches based on scattering and gathering. Each agent measures the distances from other agents and chooses its move in the direction of departing or approaching. These two approaches strengthen the discovery effect to find more undiscovered solutions and the learning effect to find more good solutions of MARTA^* respectively. We evaluate them through simulation experiments on maze and 15-puzzle problems and analyze why they work well or not from a point of heuristic depression, which is a local hollow of heuristic state evaluation values on paths to goal states. Once an agent falls into a depression, it cannot escape without filling in the depression, namely updating the state evaluation values. In a maze problem, in which deep depressions are scattered in its state space, scattered agents show better performance than gathered agents because less agents fall into depressions. On the other hand, in a 15-puzzle problem, in which shallow depressions are ubiquitous, gathered agents show better performance because they are better at cooperating to fill in the depressions. As a result, it is shown that we can get better search performance in MARTA^* by using an appropriate organization of agents depending on the characteristic of heuristic depressions of the given problem.

The 8-puzzle has been one of the popular workbenches in the artificial intelligence research. This paper analyzes the problem space of this puzzle to better understand the relationship among many problem instances. Optimal solutions of many problems can be easily obtained without search by transforming optimal solutions of one original problem. By making use of this result, all optimal solutions of all possible problems are efficiently calculated. Accurate distributions of solution lengths, number of solutions per problem and estimation error of the Manhattan distance are shown together with several important problem instances with maximum solution length(31),maximum number of solutions(86),maximum estimation error(18),etc.