An Instructable Production System; Basic Design Issues.

The full advantages of the incremental properties of production systems have yet to be exploited on a large scale. A promising vehicle for this is the task of instructing a system to solve problems in a complex domain. For this, it is important to express the instruction in a language similar to natural language and without detailed knowledge of the inner structure of the system. Instruction and close interaction with the system as it behaves are preferred over a longer feedback loop with more independent learning by the system. The domain is initially an abstract job shop. The beginning system has capabilities for solving problems, processing language building productions, and interacting with the task environment. All parts of the system are subject to instruction. The main problem-solving strategy, which permeates all four system components, is based on means-ends analysis and goal-subgoal search. This is coupled with an explicit representation of control knowledge. The systems behavior so far is restricted to simple environmental manipulations, a number of which must be taught before more complex tasks can be done. Author