This paper proposes an inductive learning environment which supports learners' knowledge acquisition from trial and error in experiments. The environment aims to assist learners in finding causal rules of experiments together with its meta-process of findings. We introduced two principal functions of the support environment : the internal experiment and the bi-directional interface. The internal experimental module manipulates the learning environment in the same way as the learner does. It tries to derive the causal rules and verifies the correctness of hypotheses by experiments. The bi-directional interface provides both the learner and the system with a channel to access the learning environment on equal condition. Both of the learner and the system has the same right to manipulate and receives the same information from the learning environment. These two principal functions provide the system with an ability to acquire rules by experiment, to monitor learner's operations and to do cooperative work with the learner. We introduced strategy space of operation and hypothesis space, so that the internal experimental module may deduce correct rules by using these two sources of knowledge, and may infer learner's state of trials in discovery. Finally, we propose methods of assisting learners together with examples excerpted from a courseware about the proportional relation.

This paper proposes a generic system architecture for microworlds that interact with the external real world. It is an extension of our microworld authoring system developed upon the IntelligentPad system. This previous system provides not only tools and objects that its users can easily combine, but also their construction kits, which enable the users to customize or to decompose the given tools and objects, or further to invent new tools and objects. Besides, a user can easily expand his microworld by importing new tools and objects from any other different microworlds. The present system has extended this previous system by introducing a new category of generic pads called proxy pads. They work as proxies of some external objects such as programs running on different machines, or computer controlled electronic or mechanical devices. These proxy pads can represent external objects, especially those in the real world. They can monitor the state and control its changes. This paper proposes the use of proxy pads for real world objects in combination with video pads that monitor their behaviors. This allows users to directly manipulate these objects through their proxy pads and observe their reaction through their video pads as if they are also members in the microworld.

Explanation can be viewed as a task of making a hearer understand the contents of an explanation produced by an explainer. This requires the explainer to take into consideration the cognitive load which the hearer will bear in understanding the explanation. Current explanation technology has focussed on reducing the load. This load reduction contributes to facilitating the hearer's understanding. However, this approach often fails to reinforce his/her understanding. Another approach to explanation is to apply a load to a hearer's understanding process. Such a load application leads the hearer to pay more attention to his/her understanding process because he/she finds it more challenging to try a heavier load. Moreover, a hearer may reach impasses due to the applied load more often than not. However, the supporting explanations for the impasses may promote the understanding process. This paper describes a load application framework, which brings these effects of the load application, and its evaluation. In the framework, understanding an explanation is regarded as building up a knowledge structure by relating the contents of the explanation to already retained knowledge. In addition, the cognitive load is represented as the amount of knowledge-structuring processes. This framework provides the following load application manner. The knowledge structure,which a hearer will finally build up, is first set up. Second an explanation is presented to the hearer. In understanding this explanation. he/she will bear the load of recollecting his/her related knowledge (recollection load) and complementing the knowledge structure by combining the given explanation with his/her knowledge (complement load). This paper also describes an experiment with the load application framework. The main purpose of this experiment is to ascertain whether the knowledge-structuring with a load is more effective for retaining the information a hearer acquires than that with no load. As a result of the experiment, we ascertained that the load application reinforced the retention. In addition, we found out the following three factors Contributing to the retention : (1) the different effect of recollection load and complement load, (2) load heaviness effect. (3) impasse effect.

In the previous study of Intelligent Tutoring System, many researchers focused on the constructing method for student model which represent the knowledge acquisition aspects of student, and most system carry out teaching based on this student model. On the other hands, considering the actual teaching environment, human teacher tries to teach the student referring not only student's domain knowledge acquisition level but also his internal psychological state, which is the "motivation level". In this paper, we focus on this student's "motivation level" in his learning situation, and we propose the new concept of ITS framework. The aim of this system is motivating the student in learning to use the system. It is important to represent the student's internal psychological state. Therefore, we propose the Human Model which are consist of Fuzzy if-then rule. In general, it is difficult to identify the fuzzy if-then rule and membership function, so we introduce the automatic fuzzy rule acquisition system, called FREGA (Fuzzy Decision Tree Generator based on Genetic Algorithm [Kunisa 93]), which is our new knowledge acquisition system mixed ID3 method [Quinlan 84] and Genetic Algorithm (GA). In following, we explain the motivation theory and propose the Human Model. Next, we show the motivation system as the ergonomic design of ITS framework and propose the human model and FREGA system. Finally, we give the estimation of this system.

Auxiliary problems are very useful to assist students in problem solving. To utilize the auxiliary problems in ITS (Intelligent Tutoring System), a framework for formulation and categorization of them are indispensable. In this paper, an auxiliary problem is defined as a simplifid problem from original one by dividing or specializing its problem solving process. First, we propose a framework of problem description which consists of surface structure, solution structure and constraint structure. Surface structure describes surface features of a problem with objects, their configuration and each object's attributes given or required. This represents a problem formulation process. Solution structure is described by a sequence of operational relations which computes required attributes from given attributes. This represents a computation process. A network composed of operational relations among attributes included in the situation specific to the problem, is called constraint structure. A solution structure is part of a constraint structure. The auxiliary problems are defined and categorized based on this description and the two simplifying operations. In this categorization, there are three types of auxiliary problems, (1) auxiliary problems with divided surface structure, which are simplified in the formulation process, (2) auxiliary problems with divided solution structure, which are simplified in the computation process,and (3) auxiliary problems with specialized constraint structure, which are simplified in the situation. We also present two experiments for justification of this definition in mechanics problems. One is the generation experiment of auxiliary problems by persons who have experience of tutoring, and the other is analysis of the problems in exercise books which seem to be used for assisting in solving another problem. Because most of the problems are adequately categorized based on our definition, it contributes to utilize the auxiliary problems in ITS.

CORES (Contextual and Relative Existence System) has been developed into a natural language processing system which can resolve pronouns and infer ellipses using its three characteristic indices : affinity nodes, temporal proximity nodes, and psychological distances. This study is intended to enable CORES predictions, playing an important part in human information processing. Contexts and Japanese particles were supposed to be influential factors in predictions and proved to be so by a priming experiment. The process of prediction were then modeled and introduced into CORES, realizing the effects of contexts on prediction by activation propagation through affinity and temporal proximity nodes and those of Japanese particles by activation levels of temporal proximity nodes. Then the validity of predictions by CORES was confirmed by comparison between the simulation and the experimental results. Finally, the problems on the attributes of verbs and particles were discussed.

In this paper, we describe a method to automatically extract Japanese auxiliary phrases from a corpus. The auxiliary phrase is a kind of idiomatic expression corresponding to auxiliary verb or postpositional particle. Typical examples are "にかんして" and "なければならない". Generally it is advantageous to handle the auxiliary phrase as one word. Therefore, building a dictionary, we need bring together auxiliary phrases like standard words. However, it is difficult to pick up auxiliary phrases. Because it is unclear to distinguish them from normal phrases. Thoroughly investigating the difference, it is defined by subjectivity of system developer. Therefore, it needs vast time to select auxiliary phrases, and there must be considerable doubt that phrases collected comprise all necessary phrases, and have uniformity. To overcome this problem, we present this method. The point of our method is to utilize the following heuristics that a auxiliary phrase has : (H1) The auxiliary phrase is consist of HIRAGANA characters. Even if KANJI character is found in it, its length is 1. (H2) Characters in front and behind of the auxiliary phrase are a certain confined characters. (H3) Each word composed the auxiliary phrase are strongly connected. Firstly, we pick up all phrases whose length is N from the corpus, however, the phrase is consist of HIRAGANA characters and KANJI characters whose length are 1. For all N(≥4), we carry out above operation. In view of (H1), all auxiliary phrases must exist in the set of phrases acquired by these operations. Then, using (H2) and (H3), we remove not auxiliary phrases from this set. Last, we remove duplicate phrases by investigating whether there is a longer phrase included the phrase. As the result, we can acquire phrases to aim in this paper. This method has a merit to easily carry out under poor environment. We made experiment on this method with ASAHI newspaper articles for one month (about 9 Mbyte). We report this result, too.

One of the main causes which cause novices to produce bugs and to reach impasses while programming is the gap between concepts used in the daily world and those used in the programming world. How to understand novices' mistakes and how to teach them are very important for an ITS. Although a lot of efforts have been devoted to the research on ITSs for novice programmers, less attention has been paid to the gap and few efforts have been made to understand novices' mistakes through the correlation between the daily world and the programming world. Our major purpose in this research is to help novices bridge the gap through the correlation between the daily world and the programming world as early as possible. The knowledge structure is organized in the three-layer hierarchy : a program model, an abstract model and a task (requirement, specification) model for better communication between the system and learners and better understanding of novices' errors. The abstract model in the intermediate layer which manages the relationship between the program model and the task model. Through these three models, the novices' bugs and misconcepts are analyzed. Bugs are understood as incorrect choices of devices in models. The reasons why novices commit errors are captured as the misconceptions about the relation of the corresponding devices in different models or misconceptions about devices of each model. Based on them, appropriate tutoring strategies are adopted.

This paper describes an adaptive model-based diagnostic mechanism. Although model-based systems are more robust than heuristic-based expert systems, they generally require more computation time. Time consumption can be significantly reduced by using a hierarchical model scheme, which presents views of the device at several different levels of detail. We argue that in order to employ hierarchical models effectively, it is necessary to make economically rational choices concerning the trade-off between the cost of a diagnosis and its precision. The mechanism presented here makes these choices using a model Diagnosability Criterion which estimates how much information could be gained by using a candidate model. It takes into account several important parameters, including the level of diagnosis precision required by the user, the computational resources available, the cost of observations, and the phase of the diagnosis. Experimental results demonstrate the effectiveness of the proposed mechanism.

Reinforcement learning aims to adapt a system to an unkown environment according to rewards. There are two issues to handle delayed reward and uncertainty. Q-learning is a representative reinforcement learning method. It is used by many works since it can learn the optimum policy. However, Q-learning needs numerous trials to converge to the optimum policy. If target environments can be described in a Markov decision process, we can identify them from statistics of sensor-action pairs. When we build the correct environment model, we can derive the optimum policy with policy Iteration Algorithm. Therefore, we can construct the optimum policy through identifying environments efficiently. In this paper, we separate learning process into two phases ; identifying an environment and determining the optimum policy. We propose k-Certainty Exploration Method for identifying an environment. After that, the optimum policy is determined by Policy Iteration Algorithm. We call a rule is k-Certainty if and only if the number of selecting it is larger than k. k-Certainty Explolation Method suppresses any loop of rules that already achieve k-Ceratinty. We show its effect by comparing with Q-learning in two experiments. 0ne is under maze environment of Dyna, the other is the environment where the optimum policy varies according to a parameter.

Given two tours of n cities, a pair of subtours of these tours consisting of the same set of cities is called a common subtour. The operation of subtour exchange crossover used in the genetic algorithm of Yamamura, et al. [山村92] is based on such common subtours. In this note, we present an O(n^2) time algorithm to enumerate all common subtours, and show that the expected number of common subtours for two random tours is at most 4+o(1).