A new paradigm called Bi-Modus Learning Environment (BLE) is introduced to support students' discovery learning. BLE is composed of the inductive learning part and the deductive larning part. The fundamental function of the inductive learning part is based on "Bidirectional Graphical Interface", "Internal Experimantation" and "Meta-knowledge for discovery learning". The paper suggests that the human discovery learning will be supported by an amalgamation of an Interactive Learning Environment like a micro world with direct manipulation into an Intelligent Tutoring System and that the central problem to realize the amalgamation is a discovery learning by a machine itself.

We consider the polynomial time inferability of primitive Prologs from positive facts. The class of primitive Prologs is a proper subclass of that of linear Prologs which is known to be inferable from only positive facts. In this paper, we discuss the polynomial time inferability of the subclass using minimal multiple generalizations. The minimal multiple generalization is a natural extension of the least generalization given by Plotkin in 1970. The minimal multiple generalization generalizes given first order words by several words, while the least generalization does by a single word. The property of the minimal multiple generalization makes it possible to perform fine generalization and to construct the heads of several clauses in a target program at the same time. We give an outline of a consistent polynomial time inference algorithm which identifies the class of primitive Prologs in the limit. The algorithm infers the heads of clauses in a target program as a minimal multiple generalization of a set of given positive facts. Furthermore, we give a similar result on the inferability of a subclass of context-free transformations which includes several well-known Prolog programs.

This paper stresses the effectiveness of stochastic approach for analyzing genetic information such as DNA sequences and protein sequences. AI technologies, especially machine learning technologies, are very attractive to extract valuable information from the enormous amounts of raw genetic information generated by biologists. To achieve this, however, more flexible and robust learning methodologies are required to deal with divergence occurring on the genetic information. In this paper, we show how stochastic approach including stochastic knowledge representations and stochastic learning algorithms works for knowledge discovery from genetic information using a motif system as an example. The motif extraction system aims to extract stable common patterns (motifs) conserved in some protein category. In the system, motifs are regarded as stochastic rules (stochastic motifs) and a genetic algorithm with Rissanen's minimum description length (MDL) principle is used as a learning algorithm. The MDL principle enables us to select "good stochastic motifs" from the viewpoint of balancing the complexity of motif and its fitness to training data. This paper also mentions about the experience of extracting stochastic motifs from super families in protein data base (PIR), the comparison of the MDL principle and the maximum likelihood method in terms of genetic algorithms, and Hidden Markov Model (HMM) representation of stochastic motifs.

Conventional knowledge-based plant control systems control a plant according to heuristics stored in a knowledge base. So, it will fail if the plant falls in the situation not considered a priori by the designer of the plant control system. We call it an unforeseen situation. This paper proposes a new architecture that can dynamically generate a plan of plant operations against the unforeseen situation. This proposed architecture is based on the thinking process of a skilled human operator. It consists of three functions as follows : (a) the diagnosis of the unforeseen situation by using qualitative causal relations among plant process parameters ; (b) the plan generation of plant operations to recover the plant from the unforeseen situation by using the constraints satisfaction method ; (c) the plan verification by predicting a plant behavior when a plant is operated according to the generated plan. To realize these functions, the proposed architecture uses the multiple models, namely, a qualitative causal model, a device model, an operation principle model and a dynamics model. We have implemented a mode-based plant control expert system on the basis of this proposed architecture. This paper focuses on the plan generation and verification of plant operations and discusses the experimental results of this expert system.

We present a method for teaching pick-and-place operation commands, generated from human assembly operation in a teaching environment, that allows a robot with mechanical constraints to perform an assembly task. The "teaching by showing" method, which is based on observing the operations of human workers, has recently been proposed for automatically generating robot commands. This method allows an operator with no knowledge of robots to teach movements to a robot. It is possible, however, that the difference between the degree of freedom of the human operator and of the robot, or differences between the teaching environment layout and the task environment layout may prevent the robot from completing the operation according to the taught sequence. In assembly tasks, the command execution order affects the completion of the assembly task. The method presented here represents operation priorities in the form of a graph, which is automatically generated from recognized human movement operations. When some operation cannot be performed by the robot, the operation sequence is reordered to complete the assembly task, on the condition that the graph remains the same. An example of a robot system is given and the results of a block assembly task are discussed.

A case-based reasoning (CBR) system solves new problems by adapting relevant cases from a case library (a case-base). When we use CBR for practical applications, we usually meet several trade-offs between CBR systems and conventional ones. In practice, the complex integration of case-based reasoners and conventional problem solvers may cause bad performance of both problem solving processes and learning processes. In this paper, we give critical analyses on the nature of CBR as a speed-up and memory-based learner, and propose quantitative measures to evaluate the performance of a CBR application system. The measures are Space Partition Ratio (SPR), Analogical Adaptation Ratio (AAF), and Reusable Case Ratio (RCR). SPR shows the utility of indices of stored cases. AAF shows the performance of adaptation of stored cases to a new problem. RCR shows direct usability of stored cases to a new problem. To validate the effectiveness of the measures, we have carried out intensive experiments on a system : IRS-CBR (Intelligent Information Retriever with a Case-Based Reasoner), which adapts the CBR method to the task of information retrieval for financial statistical databases. From the experimental results, we conclude that (1) IRS-CBR has the advantage for improving the performance of conventional information retrieval systems, (2) IRS-CBR has succeeded in both speed-up and memory-based learning in a practical sense, and (3) the proposed measures are useful for the evaluation of a CBR application.

In many problem solving systems based on generate-and-test paradigms, most of the knowledge is owned by the tester. This imbalance of knowldege distribution causes a lot of problems, such as the occurrence of useless inferences, increase in overhead and so on, even though the ATMS is employed to realize efficient problem solving. The knowledge owned by the tester cannot be applied for the scheduler immediately in all the problem solvers based on generate-and-test architecture. However, once the knowledge is cached by the ATMS, it is easy to convert the knowledge into more useful one for the scheduler. We can also obtain directly applicable knowledge for the scheduler by compiling the knowledge owned by the tester. Using this method, we can decrease the interaction and overhead between scheduler and ATMS. In practice, we can consider that ATMS have no overhead. Furthermore, we can reconstruct the scheduler with the compiled knowledge, to embed the declarative knowledge owned by the tester into the scheduler as the procedural knowledge. This paper describes three kinds of augmentation of ATMS-based problem solving systems together with evaluation of the methods.

The main aim of this research is to establish a sophisticated methodology for building expert systems based on shared and reusable large knowledge bases. Multis, one of the major Conponents of the methodology, performs task analysis interview and synthesizes problem solving engines for a given task. To design Multis the authors identify libraries of task ontology and reusable software artifacts for construction of knowledge-based systems and make them available for Synthesis via direct-interactive mapping to task models. This library consists of a set of highly generalized software primitives abstracted from existing knowledge-based systems. The mapping to the target task model is accomplished through an intermediate step in which task performers identify the correspondence of the software primitives to their own task ontology. The task ontology itself is created with the use of non-functional task primitives in the form of generic vocabulary, i. e. a vocabulary that is dependent on the task, but not the domain of expertise. The vocabulary combines into verb/noun phrases forming generic processes which are generalized conceptual primitives for a given task. In this paper one such library of software artifacts is presented for the task of scheduling (eg. classroom scheduling for an educational institution) along with the corresponding generic vocabulary and generic process library.

An intelligent tutoring system should have a student model which describes the student's understanding in order to realize adaptive tutoring. We have developed a basic architecture of HSMIS, which is an assumption-basd inductive student model inference engine. Although the algorithm of HSMIS is domain independent and is based on a complete logical foundation, the behavior of the system lacks flexibility and educational appropriateness. In this paper, we investigate which mechanisms and what knowledge is required to make HSMIS flexible and powerful. To this end, we developed the following two mechanisms : 1. Flexible decision making on the usage of information obtained from student's problem solving process. 2. Sophisticated model inference mechanism for coping with various assumptions. The control mechanism and knowledge for controlling student model inference is described in detail.

There have been many works to embed the learning capability into computers. But these researches have not yet arrived at the goal. Traditional artificial intelligence systems have been responded to only one kind of external stimulus by given knowledges but has not been able to enhance its ability or efficiency fitted to their environment. In contrast, recent language acquisition systems which efficiently learn the vocabulary and it's meanings by using linguistic and non-linguistic information has been studied. But these systems handle non-linguistic information as predicate calculus instead of the natural stimulus (as visual information) and only small categories of non-linguistic input. From this point of view, we made a concept acquisition system for the purpose of formalizing the method to acquire the concept with two external stimuli, that is, visual scene and auditory sound. Our system acquires concepts without a priori knowledge. This system learns the concepts which contain names, locations, colors and sizes of objects, using visual (image) information and the related auditory (voice) information. The basic operation is to extract a common part or feature from two images or two speech sounds, and is to map the extracted common part of images on the extracted common part of sounds. The correspondence is refined by the generalization and specialization. Consequently, some concepts are acquired about correspondence of voice features to image features, by sequencially learning from image and voice. We have realized the first stage of human's concept acquisition process on a computer system.

The major task of qualitative analysis of systems of ordinary differential equations is to recognize the global pattern of solution curves in the phase space. In this paper, I present a flow grammar, a grammatical specification of all possible patterns of solution curves one may see in the phase space. I describe flow pattern, a semi-symbolic representation of the pattern of solution curves in the phase space and show how an important class of flow patterns can be specified by the flow grammar. I then show that the flow grammar presented in this paper can generate any flow pattern resulting from any structurally stable flow on a plane. I also describe several properties of the flow grammar related to the enumeration of patterns. In particular, I estimate the upper limit of the number of application of rewriting rules needed to derive a given flow pattern. Finally, I briefly describe how the flow grammar is used in qualitative analysis to plan, monitor, and interpret the result of numerical computation.