Control of Computational Effort through Robustness Analysis: A Study of Visual Position Estimation under Time-Critical Constraints

Abstract

This work analyzes the relationship between flexible computation and robustness analysis. We start from the assumption that computational effort must be spent only when it can realistically affect decisions. We use robustness analysis to detect when decisions are (or are not) affected by changes in the agent state. As computational effort is aimed at changing the agent state, robustness analysis is the right tool for evaluating the impact of computational effort. We present models and methods that lead to efficient robustness analysis, particularly in the domain of maximum a posteriori estimation. We demonstrate this approach in an image processing system that estimates position from outdoor images.