Model Generative Reasoning

The 80s saw an explosion in AI in systems that incorporated sophisticated knowledge representation schemes, advanced reasoning capabilities, and the aim of solving real-world problems. Much of this power came through theoretical developments in non-monotonic reasoning, better implementation tools, and steadfast support from the militaryt funding agencies. Mike Coombs (now with PSL at NMSU) and I developed MGR as one such system. Its capabilities can be summarized thus:

1. a comprehensive knowledge representation scheme called Conceptual Programming based on Sowa's conceptual graphs, but extended in various ways to be useful to describe dynamic events.
2. a reasoning methodology based on the generation of alternative coherent models from the input data (the 'facts') and a stored knowledge base of conceptaul definitions. The model generation step was abductive in nature, and models could be eliminated from consideration by new facts through a failure to be coherent.
3. an implementation in Symbolics Lisp that included a graphical editor for the CP graphs and visualization of events and their relationships in time.

A paper describing some of the early work is The MGR algorithm and its application to the generation of explanations for novel events, first published in the International Journal of Man-Machine Studies in 1987. Later work is described in e-MGR: An Architecture for Symbolic Plasticity.

Development of MGR stopped when the Berlin wall came down. Our last application was to predict the behavior of enemy troops in a hypothetical ground-based war in Europe (we built a 'public' version that predicted the behavior of the King the English civil war). The army decided that this was not a viable area for research any more, so our funding dried up. Conceptual Programming, however, continued and this story can be found there.