Aaron H. Konstam

Software Science Applied to APL

Previous attempts to apply Halsteads software metrics to APL have led to inconsistent and counter-intuitive results. This work is a further investigation into the application of software metrics to APL to try to resolve some of the inconsistency. The effect of variations in the counting rules on values calculated for the software metrics was studied. These rules were used to analyze a set of programs from a previous study. In addition, a large number of APL programs from a university environment were analyzed. Evidence is presented that verifies that APL has a higher language level than any other common programming language previously studied. Counting monadic and dyadic uses of the same APL symbol as an instance of a different operator was found to have a significant effect on the language level calculated for APL. However, decomposing derived APL functions into separate operators did not seem to have a significant effect on language level.

Group classification using a mix of genetic programming and genetic algorithms

The goal of this study was to develop an algorithm for 2-group classification by discriminant analysis using genetic programming and genetic algorithms. This algorithm was tested on both theoretical and real world data. The genetic programming was used to find the functional form of the discriminant function and the genetic algorithms were then used to determine the coefficients of the functional form that maximizes the classification of the data instances into the two groups. The algorithms worked well, giving results that were comparable to or better than competing classification algorithms. 1. I N T R O D U C T I O N TO DISCRIMINANT ANALYSIS A common problem that occurs in the sciences, in the social sciences and in business is to assign new instances from a domain to one class of mutually exclusive classes based on the observed attributes of the instance. This is a problem with great theoretical and practical significance. Discr iminant analys is is one of the methods for accomplishing this task. If there are n classes or groul~s to which the instarices can be assigned, discriminant analysis is used to classify the instances in the following way. We construct a set of n functions: gi(x) , l <_ i < n (1) where: x is a vector of characteristics of the instance. Discriminant analysis attempts to formulate these functions g~(x) such that, if a instance with attributes x belongs in the

Teaching computer science principles to liberal arts students using Scheme

The Scheme dialect of Lisp is being used as an expository notation in introductory courses for liberal arts students at Trinity University. Terminology from natural language identifying parts of speech, such as verb, noun, pronoun and adverb, is used to present Scheme syntax and semantics to non programmers. Simple working models of various computer science topics are described. Experiences from two Trinity computer science courses are presented.

APL as a lingua franca in the computer science curriculum

In the last decade computer science has been struggling to establish its independent identity, pressured on one side by those who refuse to admit the existence of any new sciences and on the other by those who see computer science as no more than the art of constructing computer programs. We who are teaching computer science are caught in the middle. We must teach our students some of the art of computer technology through programming courses, but we also must instill in them those principles of the science of computing which set it apart as a discipline in its own right. We must keep ourselves from spending all our time teaching our students to program in a variety of different languages so they can get jobs as technologists. But we must also beware of spending an inordinate amount of time on the theory without teaching programming basics.

Using genetic algorithms to generate Steiner triple systems

Steiner systems, particularly triple systems, are usually generated by mathematicians using techniques from the theory of groups and quasi-groups. When pencil-and-paper enumeration becomes infeasible, mathematicians have used computers to carry out exhaustive searches. This paper presents some results of using genetic algorithms, which do not use exhaustive search, to generate Steiner systems. A specialized mutation operator was effective in generating Steiner triple systems. Future research will focus on improving the genetic algorithm to generate higher order Steiner systems whose existence is not currently known.

The use of genetic algorithms and neural networks to investigate the Baldwin effect

Standard evolutionary theory states that learned information will not be transferred into an underlying genotype. There is, however, a hypothesis that is consistent with the belief that learned behavior somehow influences the course of evolution. This hypothesis is called the Baldwin effect and it has been shown to occur in experiments with artificial life by Hinton and Nowlan and Ackley and Littman. A analysis was done of the effects of mutation and crossover rates on a computational model of the Baldwin effect which showed that this effect is most pronounced in asexual populations with low mutation rates. It was also noticed that the learning that occurred through the Baldwin effect exhibited the punctuated equilibrium behavior that is believed to be a part of all evolution.

Optimization in a distributed processing environment using genetic algorithms with multivariate crossover

We set out to demonstrate the effectiveness of distributed genetic algorithms using multivariate crossover in optimizing a function of a sizable number of independent variables. Our results show that this algorithm has unique potential in optimizing such functions. The multivariate crossover meta-strategy, however, did not result in a singularly better performance of the algorithm than did simpler crossover strategies.

N-group classification using genetic algorithms

The goal of this study was to develop an algorithm for n-group classification by non-parametric linear discriminant analysis using genetic algorithms. This algorithm tested on both theoretical and real-world data. This approach has an advantage over parametric methods in that it does not depend on knowing or being able to predict the distribution parameters of the objects being classified. It was found that this genetic algorithm, using linear decision functions and a hierarchical binary classification, is capable of classification of instances into n groups where n is greater than two. The method not only appears to work but it also appears to be able to identify the hierarchy inherent in the classification structure.

Linear discriminant analysis using genetic algorithms

Our goal was to develop an algorithm for classification by line= diwrimm“ mt analysis using genetic slgorifhms. This algorithm W= described and tested on hth theoretid and redworld da@ as well as compared to some more classical classification algorithms. Our experiments showed that our algorithm for linear dimiminant analysis gave compamble results to other commonly used methods for solving the classification problem. It has an advantage over pamtnetic methods inthatit does notdepemd unknowing or being ableto predict the distriition parameters of the objects beiig classified. The algorithm we propose turns out to be both robust snd efficient.

Performance evaluation of Ethernet and HYPERbus local area networks using computer modeling

The local area networks Ethernet and HYPERbus were simulated using the GPSS program language. Measurements of performance were network stability, messages transmitted per unit time, and the number of transmission attempts required per message. The simulations produced these results. Both networks were stable at the normal, 90 and 100 percent loads. At the 90 and 100 percent loads, Ethernet transmitted between 90.31 and 99.16 percent of the expected number of messages per unit time. HYPERbus under the same conditions transmitted 97.52 and 103.87 percent of the expected number of messages per unit time. Ethernet required up to 13 transmission attempts for some messages, with an average of 2.061 attempts per message. HYPERbus transmitted all of its messages on the first attempt.