On modularity in integrated architectures

• Authors: Steven Minton
• Venue: ACM SIGART Bulletin
• Year: 1991

As computer scientists, most of us tend naturally to write modular systems. It is ingrained in our upbringing. The PRODIGY system (Minton et al. 1989, Carbonell et al. 1990) is an example of a highly modular system. PRODIGY has separate modules for planning, explanation-based learning, abstraction creation, experimentation and analogy, to name just a few. The system is rather traditional, in that it reflects the organization of the field of AI itself. Of course, an intelligent system need not be based on such a highly modular design. A system might be highly uniform throughout, with no obvious breakdown according to function, like some neural net models might suggest. Or a system might be very modular, but the functional breakdown might not look anything like that in the PRODIGY system, where each module corresponds to a different method for analysis or learning. In this brief paper, I speculate about the role of modularity in general intelligent systems and review the reasons why modularity might be an appropriate organizing principle.