Trouble-Shooting by Plausible Inference

The PI system has been implemented with the ability to reason in both directions. This is combined with truth maintenance, dependency directed backtracking, and time-varying contexts to permit modelling dynamic situations. Credibility is propagated in a semantic network, and the belief transfer factors can be modified by the system, unlike previous systems for inexact reasoning. I TROUBLE-SHOOTING LOGIC The PI (for Plausible Inference) system enables a user to trouble-shoot physical systems in a very general way. The trouble-shooting process requires that our user first define a physical model which represents what takes place in normal operation with everything functioning correctly. If this physical model is translatable to the representation used by PI, it can be stored in computer memory and used to guide the search for the most likely failure.-In order to make the process clearer, we shall describe a few of the many methods of reasoning employed by human beings in their trouble-shooting. These methods are the ones we can at present imitate in the PI system. Suppose we have a desired goal state defined in our physical model, and this state depends on three conditions being true to attain the goal. If we execute the process and observe that the goal state was not attained, we conclude that at least one of the three conditions on which it depended must have been false, and all are possibly false. If we then perform some test designed to verify whether one of the conditions is actually true and the test shows that it is indeed true, we conclude that at least one of the remaining two untested conditions must be false. If all but one of the remaining conditions has been eliminated from consideration by further testing, we may conclude that the single condition remaining must be the guilty party. The process of elimination just described is the one normally employed by humans, and it is this process we have implemented on the computer. Of course, the three conditions may in turn have conditions on which they depend. In that case the method just described may be applied *This paper presents the results of one phase of research performed at the Jet Propulsion Laboratory, California Institute of Technology, sponsored by the National Aeronautics and Space Administration under Contract NAS 7-100. recursively to narrow the fault down further, at least to the granularity of the conditions employed in the representation. This …