Inference in intelligent machines: Application to a thermal evaporator

We present a new method of inference applicable to robots and other intelligent machines. Inferences drawn by intelligent machines are based on measurements gathered through sensory perception. We demonstrate that the methods for managing uncertainty of meaning, which recently have been extended to a wide variety of non-human systems, generally yield qualitatively incorrect results when applied to the uncertainty of evidence available to an intelligent machine. We show that even in very simple machines, no amount of sophistication in the mathematical algorithms can compensate for incorrect assumptions about the physical model. Conversely, we also demonstrate that once the essential structure of the physical model is correctly described, classical probability theory yields simple algorithms for the evaluation of the degree of evidence as it propagates through complex inference networks, including diagnostic trees and multicausal nets. As a first application, we have derived the probability algorithms relevant to diagnosing the malfunctioning of a thermal evaporator. For this system, an inference network has been constructed and compared to an implementation based on a MYCIN-type expert system. The laboratory implementation of the system is also described.