Implemented LAI systems

We say that an implemented LAI system is an information-processing artifact that instantiates some logical system and that acquires information about its environment, derives meaningful conclusions from the acquired information, behaves on the basis of these conclusions so as to manipulate its environment (with some objective in mind), acquires new information, and returns to acquiring information regarding its environment, whereupon it continues in this cycle.We here try to begin with an uncontroversial skeleton of an implemented LAI system. We recognize, however, that some systems that would intuitively seem to qualify as such only reflect a proper part of the cycle just sketched. We ignore this complication in what follows. For elaboration of the skeleton we describe, see Pollock 1995. Given this general view, an implemented AI system, and the development and refinement of one, would seem to include five basic functions, viz.,

1. Ontological Analysis. An ontology of the domain is produced. Correspondence between this domain and formulas in the logical system is established.
2. Knowledge Acquisition. The system acquires input and categorizes that input with respect to the ontology. Today, this part of the implementation task is often done by ``injection'': Humans code in knowledge to the system in question.
3. Derivation. Meaningful conclusions are derived from the information produced by 2.
4. Action. Behavior informed by the conclusions from 3 is produced.
5. Communication. The system communicates knowledge, acquired data, and conclusions therefrom (and the derivations in question) to human users of the system and to other LAI systems. Sometimes, this communication is in part accomplished by the human designer simply inspecting the ``innards'' of the system.

For example, consider a robot tank, ROB, based upon the system . The result of the analysis function might be an ontology consisting of objects called ``tanks'', ``helicopters'', ``cars'', ``people'', ``guns'', ``machine-guns'', ``soldiers'', and ``civilians.'' It might also assert that groups of tanks form things called ``tank divisions'' and that groups of soldiers form things called ``infantry divisions''. For the acquisition function, ROB might have visual sensors that it uses to find and distinguish the objects of the ontology - tanks, people, guns, etc. - from an actual battle scene. Incarnation of the derivation function might enable ROB to conclude that a person carrying a machine gun, identified by the acquisition function, is an enemy soldier and should be shot at, provided she is not flanked by a tank division. If ROB ``sees'' a person carrying a machine-gun, and various other things are true, he may act: He may shoot at this person. Finally, the communication function is responsible for conveying ROB's knowledge. ROB might produce a natural-language description of the battle scene for consumption by a field commander. This description may include possible military responses as well their justifications.

It seems to us a profitable exercise to attempt to classify the most powerful, state-of-the-art LAI systems by way of our five categories and our presentation above of logical systems and related machinery. Two such systems are OSCAR and CASSIE (in SNePS); readers are encouraged to test the analytic power of our scheme on them - keeping mind, of course, that the logical sytem at the heart of this advanced work may not coincide with a standard logical sytem.Indeed, SNePS is based on relevance logic (Anderson & Belnap 1975), which isn't covered in the volumes reviewed herein. For a look at SNePS's underlying logical system, and related matters, see Shapiro & Rapaport 1987, Martins & Shapiro 1988. OSCAR, in turn, is based on John Pollock's promising defeasible logic. For a comprehensive study of this logic and related matters, see Pollock 1995. For a narrower but very elegant and informative look, see Pollock 1992.