Engineered Robustness by Controlled Hallucination.
Jacob Beal, Gerald Jay Sussman · 2008
Most computer programs are brittle. They may have accept-able behavior for the range of applications that they are spec-ified to work for, but they fail miserably for applications even slightly outside of that range. For example, a program may work well when its inputs are complete, but be unable to pro-duce any sensible result if any of its inputs are missing or noisy. We propose a strategy for alleviating this kind of brittleness: because many programs are used repeatedly on a very sparse but highly structured subset of the possible inputs that might be provided, we can build wrappers that fill in missing or noisy data by inducing the missing information from the other inputs. We illustrate this strategy with a wrapper that acquires con-