The AIRS classifier, based on principles derived from resource limited artificial immune systems, performs consistently well over a broad range of classification problems. This paper explores the effects of adding nonEuclidean distance measures to the basic AIRS algorithm using four well-known publicly available classification problems having various proportions of real, discrete, and nominal features.

The AIRS classifier, based on metaphors from the field of artificial immune systems, has shown itself to be an effective general purpose classifier across a broad spectrum of classification problems. This research examines the new classifier empirically, replacing one of the two likely sources of its classification power with alternative modifications. The results are slightly less effective, but not statistically significantly so. We conclude that the modifications, which are computationally somewhat more efficient, provide fast test versions of AIRS for users to experiment with. We also conclude that the chief source of classification power of AIRS must lie in its replacement and maintenance of its memory cell population.

This paper presents a new supervised learning paradigm inspired by mechanisms exhibited in immune systems. This work provides an explication of a resource limited artificial immune classification algorithm, named AIRS (Artificial Immune Recognition System), and provides results on simulated data sets to demonstrate the fundamental behavior of the algorithm.

This paper presents a new tool for supervised learning, modeled on resource limited Artificial Immune Systems. A supervised learning system, it is self-regulatory, efficient, and stable under a wide range of user-set parameters. Its performance is comparable to well-established classifiers on a variety of testbeds, including the iris data, the diabetes classification problem, the ionosphere problem, and the rock/metal classification problem for mine detection.

This research employs evolution to create large (500 hidden unit) multiple layer perceptrons which attempt to perform six-way classification in a noisy environment. The classification is of words in text, according to their part of speech. The most promising aspect of the evolutionary strategy is that of allowing the entire population to evolve briefly, then split into separate populations which evolve independently, interbreed in a merged large population, split into separately evolving populations, and so on. Although this approach can be simulated on a sequential machine, it is inherently parallel. The resulting classifiers typically sustain a diversified classification strategy, as evidenced by their confusion matrices, beyond the point at which sequential steady state evolution has converged.

This article provides a description of the major components of a system that builds and updates a knowledge base by extracting the knowledge from natural language text. the knowledge extraction is done in a domain-independent manner and does not rely on particular vocabulary or grammar constructions. the only restriction is that the input text must be technical text from some specific problem domain. an important capability of the system is that it can bootstrap itself. That is, beginning with only a description of the types of object and relationships to be stored in the knowledge base, the system can start with an empty knowledge base and build the knowledge base as it processes the text. the knowledge extraction system's success in extracting knowledge from various input texts was evaluated using scoring metrics reported by Lehnert and Sundheim [AI Mag., 12(3), 81–94 (1991)]. the initial results indicate that the knowledge extraction mechanism is both effective and independent of a particular author's writing style or a particular domain. © 1995 John Wiley & Sons, Inc.

This paper presents an approach to identifying conjuncts of coordinate conjunctions appearing in text which has been labelled with syntactic and semantic tags. The overall project of which this research is a part is also briefly discussed. The program was tested on a 10,000 word chapter of the Merck Veterinary Manual. The algorithm is deterministic and domain independent and it performs relatively well on a large real-life domain. Constructs not handled by the simple algorithm are also described in some detail.

An approach to backpropagation training of a neural network for word spotting is described. It is assumed that the network has one output unit for each keyword to be detected, and that features of the speech signal are input at fixed intervals. The goal of training is to obtain a network that emits a detection pulse at the appropriate output unit when the utterance of a keyword is completed. The authors have developed a successful backpropagation strategy which incorporates 'don't care' targets for outputs expected to be in the process of rising or falling, propagation of errors for only a subset of those times at which no detection pulse is expected, iterative refinement of the temporal placement of target outputs, and use of a super-squared error criterion. In an application of the strategy to speaker-dependent, continuous digit recognition (i.e., digit spotting with no utterances of nondigits), word-error rates of 0% and 2.5% were achieved for the training and test utterances, respectively.>

Network models of natural language grow large and sparse while failing to predict many subsequent inputs. A syntax-directed speech recognizer cannot correctly transcribe a sentence for which no network path exists. The sparsity and size of a network may be reduced by partitioning the vocabulary into primary and secondary vocabularies on the basis of word frequency. Sentences with secondary phrases replaced by a placeholder are used to build a network. Secondary phrases grouped according to which primary words immediately precede and follow them are used to build lower-level networks. The groups of phrases constitute crude grammatical categories. Preliminary study suggests the efficacy of the approach.