USE OF MAXIMUM-ENTROPY METHOD AS A METHODOLOGY FOR PROBABILISTIC REASONING

Current methodologies which use probability theory for handling uncert ainty are designed primarily for use in expert systems and require ver y specific knowledge for encapsulation. However, there are other situa tions where it would be desirable to use probabilistic reasoning but i t would not be realistic to expect the knowledge to be available in a very specific form. For example, in some knowledge domains, such as en gineering/manufacturing, knowledge exists which is not in the form req uired by these systems, and, for some purposes, such as decision suppo rt, it is quite possible for the knowledge provider to know about prim ary cause/effect relationships but not be in a position to assert that other relationships are nonexistent. The paper presents a method for reasoning in small knowledge domains which overcomes the above problem . The method invokes maximum entropy theory to estimate missing inform ation and provide advice based on the knowledge available. The method is also shown to be capable of encapsulating whatever knowledge is ava ilable. The paper concludes by looking at the heavy computational requ irements of the maximum entropy method for large knowledge domains, an d it suggests that these might be alleviated, in many practical cases, because of the sparse nature of human knowledge in any domain. This w ill be the subject of further work.