An overview of ANDES: a knowledge-based scene analysis system

The presentation and control components of the ANDES system are described along with some tests that exhibit their capabilities. A particular type of associative network is used, where predicate-calculus-like formulas and decision rules share a uniform notation. Decision rules may be used in both forward, backward, or mixed direction. Procedural knowledge may be given both as heuristics and as attribute-, predicate-, or concept-associated codes. The control component performs a series of propagation of alterations, problem solving, and image measurements steps, where the latter may be executed either under request during an inference or in a spontaneous fashion. The notion of a recursive recognition cycle is present in the model as a precompiled feature, modifiable by heuristics or control options. There is also some provision for eliminating the effects of past inferences whenever an inconsistency is detected. The test results presented verify the rule chaining process, the repetition of inferences, and the mixed initiative character of the symbolic and image-related processes