When the Optimum is also Blind: a New Perspective on Universal Optimization

Consider the following variant of the set cover problem. We are given a universe $U={1,2,...,n}$ and a collection of subsets C={S_1,...,S_m} where each S_i is a subset of U. For every element u of U we need to find a set phi(u) from C such that u belongs to phi(u). Once we construct and fix the mapping phi: U -> C, a subset X of the universe U is revealed, and we need to cover all elements from X with exactly phi(X) =

bigcup_{u \in X} phi(u). The goal is to find a mapping such that the cover phi(X) is as cheap as possible. This is an example of a universal problem where the solution has to be created before the actual instance to deal with is revealed. Such problems appear naturally in some settings when we need to optimize under uncertainty and it may be actually too expensive to begin finding a good solution once the input starts being revealed. A rich body of work was devoted to investigate such problems under the regime of worst case analysis, i.e., when we measure how good the solution is by looking at the worst-case ratio: universal solution for a given instance vs optimum solution for the same instance. As the universal solution is significantly more constrained, it is typical that such a worst-case ratio is actually quite big. One way to give a viewpoint on the problem that would be less vulnerable to such extreme worst-cases is to assume that the instance, for which we will have to create a solution, will be drawn randomly from some probability distribution. In this case one wants to minimize the expected value of the ratio: universal solution vs optimum solution. Here the bounds obtained are indeed smaller than when we compare to the worst-case ratio. But even in this case we still compare apples to oranges as no universal solution is able to construct the optimum solution for every possible instance. What if we would compare our approximate universal solution against an optimal universal solution that obeys the same rules as we do? In the talk we will see that under this viewpoint, but still in the stochastic variant, we can indeed obtain better bounds than in the expected ratio model. For the Set Cover problem --- on which the main ideas and characteristics of the model will be presented --- we will see how one can obtain H_n-approximation, which matches the approximation ratio from the classic deterministic setup. Moreover, this result holds for all possible probability distributions over U that have a polynomially large carrier, while all previous results pertained to a model in which elements were sampled independently. The result is based on rounding a proper configuration IP that captures the optimal universal solution, and using tools from submodular optimization.