Keki B. Irani

On the Handling of Continuous-Valued Attributes in Decision Tree Generation

We present a result applicable to classification learning algorithms that generate decision trees or rules using the information entropy minimization heuristic for discretizing continuous-valued attributes. The result serves to give a better understanding of the entropy measure, to point out that the behavior of the information entropy heuristic possesses desirable properties that justify its usage in a formal sense, and to improve the efficiency of evaluating continuous-valued attributes for cut value selection. Along with the formal proof, we present empirical results that demonstrate the theoretically expected reduction in evaluation effort for training data sets from real-world domains.

Improved decision trees: a generalized version of ID3

Abstract A popular and particularly efficient method for inducing classification rules from examples is Quinlans ID3 algorithm. This paper examines the problem of overspecialization in ID3. Two causes of overspecialization in ID3 are identified. An algorithm that avoids some of the inherent problems in ID3 is developed. The new algorithm, GID3, is applied to the development of an expert system for automating the Reactive Ion Etching (RIE) process in semiconductor manufacturing. Six performance measures for decision trees are defined. The GID3 algorithm is empirically shown to outperform ID3 on all performance measures considered. The improvement is gained with negligible increase in computational complexity.

Applying machine learning to semiconductor manufacturing

The generalized ID3 (GID3) algorithm, which takes a training set of experimental data and produces a decision tree that predicts the outcome of future experiments under various, more general conditions, is described. The tree can then be translated into a set of rules for an expert system. Two extensions to GID3MmRIST, and KARSM-that deal with the problems of noisy data and the limited availability of training data are discussed. The application of GID3 to reactive ion etching manufacturing process diagnosis and optimization and to knowledge acquisition for an expert system is described.<<ETX>>

Queueing network models for concurrent transaction processing in a database system

This paper presents two queueing network models which correspond to different implementations of the lock management algorithm for concurrent transaction processing in a database system. These models are developed to investigate the effects of varying the granularity of locks and the degree of multiprogramming on the performance of a database system. A numerical example is presented for a set of apparently realistic parameters and its results are discussed. In addition to other conclusions, these results also confirm the result of Ries and Stonebraker, using a simulation model [9], that a relatively coarse granularity is sufficient to allow enough parallelism for efficient resource utilization. In contrast with simulation models, the queueing network models presented in this paper allow us to examine more closely the cause-effect relationships of concurrent transaction processing in a database system at less cost.

The join algorithms on a shared-memory multiprocessor database machine

The authors develop and present a large set of parallel algorithms for implementing the join operation on a shared-memory multiprocessor database machine. The development of these algorithms follows a structured approach. The major steps involved in the processing of the join operation by the machine are first identified. Then, alternative join algorithms are constructed by concatenating the different ways of performing these steps. A study of the performance of the proposed algorithms is presented. This study shows, among other things, that for a given hardware configuration there is not just one overall best performing join algorithm, but rather different algorithms score the best performance, depending on the characteristics of the data participating in the join operation. >

Fragmenting relations horizontally using a knowledge-based approach

In distributed DBMSs, one major issue in developing a horizontal fragmentation technique is what criteria to use to guide the fragmentation. The authors propose to use, in addition to typical user queries, particular knowledge about the data itself. Use of this knowledge allows revision of typical user queries into more precise forms. The revised query expressions produce better estimations of user reference clusters to the database than the original query expressions. The estimated user reference clusters form a basis to partition relations horizontally. In the proposed approach, an ordinary many-sorted language is extended to represent the queries and knowledge compatibly. This knowledge is identified in terms of five axiom schemata. An inference procedure is developed to apply the knowledge to the queries deductively. >

Automatic data base schema design and optimization

The production of an appropriate CODASYL Data Base Task Group (DBTG) Data Description Language (DDL) schema for a given data management application is a significant design problem. This research is devoted to the development of a methodology to automate and optimize the design of DBTG schema structures, using analytic modelling and optimization techniques. Given an implementation independent description of the data management requirements, it is possible to produce a schema configuration which is optimized with respect to logical record access rate, subject to storage and feasibility constraints, within a selected class of schemas. The storage/access rate trade off is expressable as an integer program, which can be mapped into a network traversal problem with a known dynamic programming solution.

On Network Linguistics and the Conversational Design of Queueing Networks

Network diagrams are a frequent means of problem description in many technical disciplines. However, problem-oriented graphic systems which use network diagrams as the medium of communication require translators. These accept the information provided in the network diagrams, associate mathematical meaning with the symbols of the diagram, and then transform that meaning into a model of the entire network which is capable of computer solution. A formal framework for these operations, constituting a useful linguistic approach to the translator functions, is offered. The approach is illustrated for Markovian queueing networks.

Markovian queueing network models for performance analysis of a single-bus multiprocessor system

An exact solution for the performance analysis of a typical single-bus multiprocessor system is presented. The multiprocessor system is modeled by a Markovian queueing network. An r-stage hypoexponential distribution or an r-stage hyperexponential distribution is used to represent the nonexponential service times. Consequently, the equilibrium probabilities of two-dimensional Markov chains are expressed by simple recurrence relations. Processing efficiency is used as the primary measure of performance. To investigate the effects of different service time distributions on system performance, comparative results are obtained for a large set of input parameters. The numerical results illustrate that processing efficiency attains its maximum value for a constant (deterministic) service time; if service time of common memory is hypoexponentially distributed, then approximating the service time by an exponential distribution produces less than 6% error on the system performance. >

A methodology for solving problems: problem modeling and heuristic generation

A methodology is given for modeling a problem and solving it using the A* algorithm. The heuristic used for A* is mechanically generated from the simplified problem, which is derived by relaxing each of the predicate formulas describing the rules and the goal state of the problem. The generated heuristic satisfies the conditions of admissibility and monotonicity. The methodology is applicable for solving general problems. The overall procedure for this methodology is illustrated by four well-known problems, namely, the eight-puzzle problem, the traveling salesman problem, the robot planning problem, and the consistent labeling problem. The values of the heuristics generated by this procedure are compared to the corresponding values of problem-oriented heuristics reported in the literature. >