Encounters with Expert Systems

In the early eighties, I became fascinated by artificial intelligence, without entering the field seriously yet, but in 1986, having heard of the "wonders" accomplished by computer scientists with so called expert systems, I undertook and completed the systematic reading of several thousand pages of major publications in the fairly chaotic field of expert systems, artificial intelligence, computer vision.

To deepen my understanding, I launched in 1987, a yearly international autumn school on artificial intelligence, expert systems, and computer vision, co-sponsored by the industrial R&D division of the THOMSON company, and Ecole Normale Superieure. As scientific director, and proceedings editor for this event which gathered hundreds of computer scientists in France every year from 1987 to 1990, I became quite aware of the major trends in thess exciting fields, and of obvious weaknesses in their mathematical formalizations.

Simultaneously, as a scientific consultant, I was closely involved with the effective realization of two industrial experts systems at RHONE-POULENC France : one expert system, based on the G2 software of Gensym, for process control of phosphate production plants, another one based on N.Expert Object softwares, for selecting and dimensioning adequately the essential equipments of chemical plants.

This practical industrial research activity in experts systems, in the background of my mathematical research, convinced me quickly of the very strong limitations of expert systems, and of the huge gap between human intelligence and the narrow expressive power of rule based reasoning.
The promising complexity of research domains such as automatic sound and image analysis and machine emulation of human perception and learning capacities required strongĀ  conceptual jumps, and I became quite aware of the potential impact, for artificial intelligence, of barely emerging new powerful mathematical formalisms, most of them particularly exciting for massively distributed computing techniques :

- Simulated annealing and Stochastic relaxation
- Markov random fields and artificial vision
- Formal neural networks and Automatic learning.

Selected References R. Azencott :
[37] [39] [44] [50] and co-realisation of several industrial experts systems softwares