Supermodels and Robustness

When search techniques are used to solve a practical problem, the solution produced is often brittle in the sense that small execution difficulties can have an arbitrarily large effect on the viability of the solution. The AI community has responded to this difficulty by investigating the development of “robust problem solvers” that are intended to be proof against this difficulty. We argue that robustness is best cast not as a property of the problem solver, but as a property of the solution. We introduce a new class of models for a logical theory, called supermodels, that captures this idea. Supermodels guarantee that the model in question is robust, and allow us to quantify the degree to which it is so. We investigate the theoretical properties of supermodels, showing that finding supermodels is typically of the same theoretical complexity as finding models. We provide a general way to modify a logical theory so that a model of the modified theory is a supermodel of the original. Experimentally, we show that the supermodel problem exhibits phase transition behavior similar to that found in other satisfiability work.