LANSKY@SRI-WARBUCKS.ARPA (Amy Lansky) (07/23/86)
OUR COGNITIVE ABILITIES LIMIT THE POWER OF AI Jack Alpert (ALPERT@SCORE) Stanford Knowledge Integration Lab and School of Education, Stanford University 11:00 AM, MONDAY, July 28 SRI International, Building E, Room EK228 "Expert Systems: How far can they go?" was a panel topic at AAAI 1985. Brian Smith described the limits of AI in terms of the programmer's ability to know if his encoded model reflected the world that his expert system was to manage. "We have no techniques.. to study the ... relationship between model and world. We are unable... to assess the appropriateness of models, or to predict when models fail." Most of us with icy road experience are convinced we know how to recover from skids. In the talk I will prove that our skid recovery algorithms work only on a small set of possible skids. Skids that lie outside of this small set result in accidents. Our "inappropriate" skid recovery models cause accidents. 20 years of driving experience does not revile the skid model's limitations. When we have been building expert systems for 20 years, why should we be any better prepared to perceive model inappropriateness? The limited set of cognitive abilities that most people develop cannot identify domains where models fail. I describe a temporal cognitive ability most of us lack. Given the definition of such an ability, I will briefly describe a line of research that explains why people never develop the ability. Should this research be successful, we will create new learning environments that enhance first cognitive abilities, then modeling, and finally the power of AI systems. VISITORS: Please arrive 5 minutes early so that you can be escorted up from the E-building receptionist's desk. Thanks!