|Intelligent Systems And Their Societies||Walter Fritz|
PRODIGY, was developed by Jaime Carbonell, Oren Etzioni, Yolanda Gil, Robert Joseph, Craig Knoblock, Steve Minton and Manuela Veloso at the School of Computer Science, Carnegie Mellon University. It was described in the SIGART journal in August '91. PRODIGY uses objectives, situations, response rules, and plans. The response rules (called operators) have a stimulus (precondition) and a response (a list of effects). The stimulus is expressed by a system (called predicate logic) which includes negation, conjunction, disjunction, existential, and universal quantifiers.
Negation means part of a stimulus of a response rule should not be in the present situation. Conjunction means several parts should be there together. Disjunction means one or the other should be present. Existential quantifier means at least one should be present. Universal quantifier means all should be present.
Introducing Control Rules
As with many IS systems, the central part of the PRODIGY system is a planner. This planner generates a sequence of response rules (called operators) in an attempt to get from a given situation to a set of objectives. PRODIGY's planner is different from the systems we have seen so far in that its search for applicable response rules is, in turn, guided by control rules. These control rules take the form of a stimulus side -- describing the condition(s) to which it is applicable -- and a response side -- for selecting, rejecting or preferring a response rule. Control rules can be general or applicable only to the problem at hand (the specific environment) and can be inserted through its initial programming or learned by the system. These control rules are used during the search for response rules in order to deal with any decision points that arise.
The planner works by starting with only the situation and the objective. It then decides where to expand. For instance, it may choose to begin working with the current situation and one of the objectives. It then selects a response rule and applies it to produce a new intermediate situation. The process is then repeated for any of the existing situations and any of the objectives. When all the objectives are satisfied the search stops. During this process, the control rules determine where and how to work. If no control rule is applicable, the system works by chance. If the planner finds that this chance has resulted in a bad decision, it backtracks and tries to learn the control rule that must have been missing.
PRODIGY has a number of learning modules:
The system forgets control rules that are less useful.
PRODIGY has been used for robot path planning, for matrix algebra manipulation, for machine shop planning and other uses.
Copyright © New Horizons Press