This paper was originally published as Technical Report (no. 189) for the CSIS (Computer Science and Information Systems)Department in 2003.

Document Type



Artificial Neural Networks (ANNs), a class of machine learning technology based on the human nervous system, are widely used in such fields as data mining, pattern recognition, and system control. ANNs can theoretically learn any function if designed appropriately, however such design usually requires the skill of human expert. Increasingly, Genetic Algorithms (GAs), a class of optimization tools, are being utilized to automate the construction of effective ANNs. The Wisconsin Card Sorting Test (WCST) is a tool used by psychologists to assess human subjects' planning and reasoning ability. The adaptive learning required in the test's task and its ambiguous nature make it an interesting one to use as a test of the learning properties of ANNs. In this paper, an ANN model is presented that is potentially capable of learning the WCST task. The model was developed based on the division of the WCST task into three sub-tasks. Six GAs and one non-generic search algorithm were used to design two ANNs to learn two of these sub-tasks. Each learned its sub-task to a high degree of accuracy. One of the sub-tasks required a training pattern set with ambiguous input-output mappings. The nature of backpropagation learning on this pattern set was unusual in that it was non-linear. The performance of the search algorithms was compared. The results imply that local search was a more effective operator than global search for this task. A Lamarckian GA outperformed Baldwinian GAs, which in turn outperformed Darwinian GAs. A novel GA referred to as Reverse Baldwinian was also less effective than the Lamarckian GA. The NOn-Genetic algorithm performed comparably to the Lamarckian GA, in addition to being more efficient. GEneral difficulties in using GAs to evolve ANNs that have been noted in previous research may have been responsible for these results. Additionally, the suspected ease of learning both training pattern sets and the effects of the ambiguity of one of the pattern sets may have impacted the algorithm's performance.