Daniel C. St. Clair

Dynamic ID3: a symbolic learning algorithm for many-valued attribute domains

Quinlan’s ID3 machine learning algorithm induces classification trees (rules) horn a set of tiaining exsntplea. The algorithm is extremely effective when training examples are composed of attributes whose values are taken from small discrete domains. The classification accuracy of ID3-po&tced trees on domains whose attributes are many-valued tends to be margirtaf due to the large number of possible values which may be associated with each attribute. Attempts to solve this problem by a priori grouping of attribute values into distinct subsets has met with limited success.

An Inferencing Language for Automated Spatial Reasoning about Graphic Entities

A method is proposed for automated reasoning about graphic entities. First, a formal representation scheme is suggested for persistently storing graphic information as fundamental graphic entity types. Next, fundamental relationships between these types are identified. A formal, graphic entity reasoning based inference language (GERBIL) is then presented to implement the relationships. An architecture is proposed, linking a computer graphic system for persistent entity storage with a knowledge based system shell for inferencing. A prototype system, Dafne, demonstrates proof of principle.

Rule-based machine learning of spatial data concepts

Extensive work has been done on interfacing expert systems with spatial systems such as CAD (computer aided drafting) or GIS (geographic information systems). Likewise, much work has been done on the use of machine learning algorithms to mechanically build the rules which are input into expert systems. This paper explores one particular combination of these areas of research. The rule-based learning algorithm AQ15 was used to classify spatial data from a GIS. A variety of miscellaneous annotation features were classified and input into AQl5 as training data. In order to produce rules which would allow an expert system to reclassify the annotations, AQl5 would need to learn spatial concepts such as “parallel” or “close-to.” The resulting knowledge base was used to validate existing geographic data.

Formation of clusters and resolution of ordinal attributes in ID3 classification trees

Many learning systems have been designed to construct classification trees from a set of training examples. One of the most widely used approaches for constructing decision trees is the ID3 algorithm [Quinlan 1986]. Decision trees are ill-suited to handle attributes with ordinal values. Problems arise when a node representing an ordinal attribute has a branch for each value of the ordinal attribute in the training set. This is generally infeasible when the set of ordinal values is very large. Past approaches have sought to cluster large sets of ordinal values before the classification tree is constructed [Quinlan 1986; Lebowitz 1985; Michslski and Stepp 1983]. The approach presented in this paper dynamically forms sinlge-conclusion and multiple-conclusion (overlap) clusters at leaves in the decision trees. The experiments with the Lebowitz gap-finding method resulted in clusters which differentiated classes very poorly. Domain knowledge is used directly by the overlap resolution algorithms to &termine the number of clusters to form. The multiple-conclusion clusters are then resolved by using techniques [Hacke 1990]. Results indicate that overlap resolution techniques capture the ability to predict the classification of the unseen instances. The results of accuracy increase and learning-rate experiments are encouraging. Experiments are performed to determine the techniques to which maximum accuracy-increase is attributable. The numerical results of the experiments are presented in this paper.

Analysis of rule sets generated by the CN2, ID3, and multiple convergence symbolic learning methods

Since symbolic learning methods develop distinctive sets of rules when given identical training data, questions arise as to the quality of the different rule sets produced. The results of this research provide techniques for comparing and analyzing rule sets. Numerous rule sets were generated using three wellknown symbolic learning methods; Quinlan’s ID3, Clark and Niblett’s CN2, and Murray’s Multiple Convergence algorithm. The analysis techniques were then applied to evaluate these sets of rums. The techniques as well as a guide for using them are presented in a concise summary following the discussion of the experimental results.

Design techniques for the control of errors in backpropagation neural networks

A significant problem in the design and construction of an artificial neural network for function approximation is limiting the magnitude and variance of errors when the network is used in the field. Network errors can occur when the training data does not faithfully represent the required function due to noise or low sampling rates, when the networks flexibility does not match the variability of the data, or when the input data to the resultant network is noisy. This paper reports on several experiments whose purpose was to rank the relative significance of these error sources and thereby find neural network design principles for limiting the magnitude and variance of network errors.

Effects of nonsymmetric release times on rate monotonic scheduling

This paper discusses problems associated with scheduling periodic tasks on a uniprocessor in a hard, real-time processing environment using a static-priority, preemptive-resume operating system. The scheduling problems associated with a task set containing a single periodic task which has two fixed release periods of unequal length are examined. Some real-world applications may require task release times which are periodic, but whose tasking periods are not symmetric. A scheduling algorithm for task sets with a single nonsymmetric task was developed for static-priority, preemptive-resume operating systems. The nonsymmetric scheduling algorithm is based on the rate monotonic scheduling algorithm which assigns higher task priorities to tasks with shorter release periods. The effects on processor utilization using two different priority assignment schemes are examined. The first method sorts the task priorities by the average release periods. The second method sorts the task priorities using the short nonsymmetric task period with the average period lengths for the remaining tasks. A large number of task sets were generated to characterize effects of the two priority assignment methods on task set utilization levels when used with the nonsymmetric scheduling algorithm. Characterization results for the two methods indicated that the short nonsymmetric task period priority assignment had higher breakdown utilizations than the average period priority assignment method. For task sets with a low utilization nonsymmetric task, use of the short period priority assignment method resulted in little or no loss in the overall task set breakdown utilization. Maximizing the processor utilization is desirable, provided the tasks operate in a deterministic manner and meet their deadlines. The nonsymmetric scheduling algorithm allows the system designer to calculate the feasibility of a task set containing a nonsymmetric period task.

Effect of the x 2 test on construction of ID3 decision trees

Quinlan’s 1D3 machine learning algorithm induces classification trees (rules) from a set of Eaining examples. The algorithm is extremely effective when training examples contain little or no noise. Noisy training data may result in the induction of decision trees which are not representative of the domain being modeled, To reduce the effect of noise on ID3’s construction of decision trees, Quinlarr suggested the use of the Z2 statistic as a tool for identifying noise in training attributes. Attributes appearing ttr be noisy can then be excluded from consideration as branching attributes. Since an a[fribute may appear more than cmce in a tree, an attribute musl be tested for noise each time it is under consideration. The work presenled in this paper evaluates the effect of using the Z* statistic as a tool for identifying noise during ID3 tree construction. [t was found that the effects of this approach vary Jepending on the classifl( wiur. criterion used to match tree and test example conclusions. Numerical results are provided which illustrate these differences.

ESPAD: an adaptively controlled rule-based expert system for monitoring and diagnosing space vehicle subsystems

The Expert System for Power Subsystem Analyses and Diagnoses (ESPAD) provides an architecture for designing embedded rule-based expert systems whose function is to monitor and diagnose space vehicle subsystems. It introduces the concept of adaptive control structures which use component failure histories to expedite expert system response. Although ESPAD specifically monitors and diagnoses a subsystem for converting water to H2 and O2, the approach is applicable to a large number of subsystems which permeate space vehicles.

Are the university computer sciences satisfying industry (Panel Discussion)

B a r r y B. F l a c h s b a r t M c D o n a l d-D o u g l a s A u t o m a t i o n Co.