IRA@WASHINGTON.ARPA (07/05/83)
From: Ira Kalet <IRA@WASHINGTON.ARPA> This is in response to the query about when to build an AI "front-end" to an existing software system as a separate process with its own address space, as opposed to putting more code in the existing system to implement the AI component. At the University of Washington we have built a very complex graphic simulation system for planning of radiation therapy treatments for cancer. We are now starting to work on a rule based expert system that will model the clinical decision making part of the process, with the two (separate) systems to communicate via messages. We do this as two separate processes because the simulation system is already a system of multiple concurrent processes communicating by messages, and because the simulation system is written in PASCAL, which seems less suitable than, for example, INTERLISP, for the AI component. The kind of information needed to pass between the systems also affects the decision. In our case, the AI system will consult the graphic treatment planning system for answers to questions that are rather traditionally compute intensive, eg. radiation dose calculation, geometric calculations...so the messages are simple and well defined.