The purpose of this paper is to develop a new method for constructing Bayesian belief networks. The unique feature of this study is that constructing Bayesian belief network process is regarded as decision making process. In terms of the expected utility introduced from this view point, while adding arcs procedure does not attain the maximization of the expected utility, the reducing arcs procedure, which is introduced from the maximization of expected utility, attains it. We performed some stochastic simulation experiments, the results of which show that this procedure found an optimum structure.

Recently, electronic mail has been widely used for a variety of purposes, such as cooperative works, advertisement, and informal communication. In order to support such communication forms, the conventional mailing-list systems cannot offer flexibility in management of group members, such as addition, update, and deletion of members. Furthermore, even if we employ another approach where user information is managed for each individual and groups are constructed dynamically according to the specification of destination address condition, all information must be described in a predefined format, which prevents us from flexibly describing the way of information distribution. In this paper, we propose a method for dynamically constructing groups using a reasoning mechanism and discuss the design and implementation of MILD (MaIL Distribution) system, which is based on the proposing group construction method. In MILD, we can describe various information regarding, e.g., individuals, business, and hobbies, and such information can be intuitively integrated without explicitly specifying how to integrate them. Furthermore, providing the description capability of operations on the results of reasoning, various destination addresses can be specified by combining these operations. As a result, we show that MILD is flexible enough for both on the user information management and on the specification of conditions.

We have proposed the SIP planning method which interleaves planning with executions by using the success probabilities of plans. With SIP, an agent is able to control deliberation depending on the change of an environment. The formal framework has been developed using a plan Bayesian network, however we need to investigate the utility of SIP in a testbed for a dynamic environment. In this paper, we report results for various experiments made in a simplified Tile World. First we characterize a simplified Tile World with a few essential parameters: dynamics, cost for actions including observations and uncertainty, Various experiments are made, and we find out interesting properties: a single optimal execution threshold between reactivity and deliberation, robustness of the optimal threshold against the change of dynamics and observation costs. We consider that these properties significantly contribute to design of an agent in a dynamic environment.

Information flow in a dynamic environment, such as a production process and a communication network, is a valuable resource for understanding the general behavior of the environment, discovering the regularities and anomalies currently happening in the environment, controlling the evolution process, and modeling and intelligent management of the environment. Unfortunately, the data generated in a dynamic environment are often expressed in low level primitives and in huge volumes. Because of the dynamic, continuous and rapid changes of the information flow, it is difficult to catch the regularities and anomalies in a dynamic environment and react promptly for real-time applications. In this paper, a knowledge discovery technique is integrated with data sampling and active database techniques in order to discover interesting behaviors of a dynamic environment and react intelligently to the environment changes. The discovery of the dynamics in a computer communication network and the application of the discovered knowledge for network management are taken as an example in our study. The study shows (1) data sampling is necessary in the collection of information for regularity analysis and anomaly detection; (2) knowledge discovery is important for generalizing low level data to high-level information and detecting interesting patterns; (3) active database technology is essential for real-time reaction to the changes in a dynamic environment; and (4) an integration of the three technologies forms a powerful tool for control and management of large dynamic environments in many applications.

Many conventional works in reinforcement learning are limited to Markov decision processes (MDPs). However, real world decision tasks are essentially non-Markovian. In this paper, we consider reinforcement learning in partially observable MDPs(POMDPs) that is a class of non-Markovian decision problems. In POMDPs assumption, the environment is MDP, but an agent has restricted access to state information. Instead, the agent receives observation containing some information about states of the MDP. Also we focus on a learnig algorithm for memory-less stochastic policies that map the immediate observation of the agent into actions: The memory-less approaches are suited for on-line and real-time adaptive systems that have limited memory and computational resources. Then, the following mathematical results are got. First, it can improve its policy to maximize immediate reward by stochastic gradient ascent without estimating any state or immediate reward. Second, it can improve the policy to maximize discounted reward in an initial state by stochastic gradient ascent without estimating any state, immediate reward or discounted reward. The above advantages are remarkably effective in POMDPs, because it is not required to estimate any states, immediate reward or discounted reward explicitly. Making use of these results, we present an incremental policy improvement algorithm to maximize the average reward in POMDPs. We ensure the rational behavior of the proposed algorithm in a simple experiment.

The attributes to explain features of numeric dataset have been extracted depending on the similarity based on the distance between all pairs of data. The new extraction method was developed by paying attention to the symmetric relation over the entire data range that was called the relational cohesiveness in this article. This method was applied to investigate for the properties of normal alchohols, and it was found that some chemical features including the discrimination rule of the normal chain compounds were extracted by using both this method and the concept of the valence of the carbon and hydrogen. The method developed here can be applied to derive the knowledge from any numeric dataset in the database and measurement data, etc.

The construction of decision trees by the ID3 algorithm is a very well known approach to inductive learning. However, it is necessary to prune decision trees constructed by the ID3 algorithm, because it cannot deal well with uncertainty due to noise in the data. When redundant features are included in a given feature set, the ID3 often tend to generate over-specialized decision trees. In recent years, the feature selection problem has been closed up in machine learning. This paper emphasizes that the feature selection as pre-processing, the complete tree generation as central-processing, and the pruning as post-processing should be unified. In general, accuracy and simplicity are requested of a decision tree. There is a tradeoff relation between accuracy and simplicity. This paper emphasizes that the decision tree induction should be formulated as a multi-objective optimization problem. The rational solutions of such a problem are known as Pareto optimal. This paper presents a genetic algorithm for generating Pareto optimal decision trees at once. The fitness is defined as a vector function of minimizing the error rate and minimizing the number of leaf nodes. The sub-trees exchange crossover and a sub-tree insertion as mutation are adopted to generate new decision trees. The non-Pareto optimal selection strategy is introduced as a model of generation alternation. Under this strategy, the population can come near to the true Pareto optimal set in progression. The algorithm is applied to the Digit benchmark problem and compared with the traditional approaches. The experiments show that the proposed algorithm can generate Pareto optimal solutions that dominate completely solutions obtained by the existing methods.

Recently, Gelfond and Lifschitz proposed an efficient computational method for prioritized circumscription which compiles circumscriptive theories into logic programs. But there are such difficulties in their method that a given circumscriptive theory cannot always be compiled successfully due to strong assumption about the syntax of a circumscriptive theory. So, we extend their method in order to expand its applicable class while keeping its computational efficiency. Our idea is to transform a given circumscription into a logically equivalent one in which such difficulties disappear. In this paper, we show it can be done by making use of Lifschitz's result that some parallel circumscription can be replaced by an equivalent first-order theory. As a result, some class of prioritized circumscription, which cannot be handled by Gelfond and Lifschitz"s method, can be compiled into logic programs by our method.

This paper describes a novel method for preferring hypotheses in hypothetical reasoning. Traditional studies on the preference have focused on the way to select the best hypothesis with a syntactic criteria: the minimum explanation is best. Unfortunately the criteria does not contain the cost for observing the environment in order to verify the selected hypothesis. However, considering an agent which observes and recognizes the environment with hypothetical reasoning, we can not ignore the cost, utility and uncertainty for observing the environment. Therefore we propose the preference using cost, uncertainty and utility of observing and verifying hypothesis. We argue our method is more suitable for an agent acting the environment than syntactic criteria, and evaluate the performance through the experiments in pattern recognition.

KICK-SHOTGAN is known as the fast hypothetical reasoning system which avoids inefficient backtracking by forming a compiled inference-path network followed by the foward synthesis of necessary hypothesis combination along this network. In this paper, we propose an improvement of the KICK-SHOTGAN"s consistency check on the synthesis of some hypotheses using inference-path network.

k-Certainty Exploration Method gives top priority for selection to an action whose number of selection is the fewest. However it doesn't consider any state-transition probability. Therefore, though it guarantees the rationality and the efficiency under deterministic Markov decision processes (MDPs), it doesn't always guarantee the rationality nor the efficiency under stochastic MDPs. In this paper, we propose l-Certainty Exploration Method which is an extension of k-Certainty Exploration Method to stochastic MDPs. We define reachability as a difference between the sampled number of any rule and the sampling number necessary to identify its rule structure by error e with confidence 1. l-Certainty Exploration Method realizes efficient identification of the environment through selecting prior to a rule whose reachability is the lowest. We show the superiority of l-Certainty Exploration Method compared with the other methods through an numerical example.