This paper examines the three prevalent approaches to Artificial Intelligence (AI), namely symbolic reasoning systems, connectionist systems, and emergent systems based on the principles of the subsumption theory. Distinguished by their top-down and bottom-up mechanisms all three approaches have strengths and weaknesses. While the logical reasoning approach is precise and well supported by mathematical theories and procedures, it is constrained by a largely predefined representational model. Connectionist systems, on the other hand, are able to recognize patterns even if these patterns are only similar and not identical to the patterns that they have been trained to recognize, but they have no understanding of the meaning of those patterns. The subsumption approach appears to overcome many of the weaknesses of the other two approaches in theory, but there is concern that it may not scale to more complex real world applications.

The author points out that in addition there are weaknesses that all three AI approaches share, namely inability to deal with exceptions, lack of mechanisms for analogous comparisons, and very primitive conceptualization capabilities at best. It is noted that the human agent performs decidedly better in these areas.

The paper concludes with the proposition that only a hybrid approach holds sufficient promise to meet the full expectations of intelligent systems. It is further suggested that this hybrid approach should include the contributions of the human agent as an integral component of the intelligent system, in most cases.


Software Engineering



URL: http://digitalcommons.calpoly.edu/cadrc/43