Abstract
Conventional approaches to most image understanding problems suffer from fragility when applied to natural environments, where the complexity in the environment becomes overwhelming. Complexity in Intelligent Systems can be managed by breaking the world into manageable contexts. GRAVA is a re.ective architecture that supports self-adaptation and has been successfully applied to a number of visual interpretation domains. The GRAVA architecture supports robust performance by treating changes in the program?s environment as context changes. Automatically tracking changes in the environment and making corresponding changes in the running program allows the program to operate robustly. We describe the architecture and explain how it achieves robustness. In particular, we present an algorithm based on Minimal Description Length (MDL) that permits contexts to be automatically induced from corpus training data. The algorithm does not require prior assignment of the number of contexts.