Catalogue of Artificial Intelligence Techniques
Keywords: Explanation-Based Generalisation, learning, preconditions
Author(s): Bernard Silver
Precondition analysis is an analytic strategy-learning technique. Precondition analysis operates in two phases: the learning cycle and the performance cycle. In the learning cycle, the program is given an example of a correctly executed task. The example may contain several individual steps, each step being the application of an Operator. The program first examines the example, to find out which operators were used in performing the task. This stage is called Operator Identification. During this phase, the program may discover that it doesn't possess the relevant operator, and the user is asked to provide the necessary information. Once this phase is complete the program builds an explanation of then strategic reasons for each step of the task. The explanation is in terms of satisfying the preconditions of following steps. From this explanation, it builds a plan that is used by the performance element. These plans are called schemas. The performance element executes the schemas in a flexible way, using the explanation to guide it. The explanations are used to make sensible patches if the plan can't be used directly. Precondition analysis has been implemented in LP (Learning PRESS), a program that learns new techniques for solving algebraic equations. Precondition analysis is somewhat similar to the explanation-based generalisation (see Explanation-based Learning) approach of Mitchell, but differs in that precondition analysis can work in domains in which the methods are not invertible, and in situations where the domain theory is imperfect.
- Silver, B., Precondition analysis: learning control information, Machine Learning: An Artificial Intelligence Approach (Michalski, R.S., Carbonell, J.G. and Mitchell, T.M.
, eds.), vol.2, Morgan Kaufmann, Los Altos, California, 1986, pp.647--670 (Chapter 22).