Online Problems and Two-Player Games: Algorithms and Analysis

Abstract

In this thesis we study three problems that are adversarial in nature. Such problems can be viewed as a game between an algorithm and an adversary, where the adversary always tries to force the algorithm into worst-case scenarios during its execution. Many real world problems with inherent uncertainty or lack of information fit into this model. For instance, it includes the vast field of online problems where the input is only partially available and an adversary reveals the complete input gradually over time (online fashion). The algorithm has to perform efficiently under this uncertainty. In contrast to the online setting, in an offline setting, the complete input is available in the beginning. The first problem that we investigate is a classical online scheduling problem where a sequence of jobs that arrive online have to be assigned to a set of identical machines with the objective of minimizing the maximum load. We study a natural generalization of this problem where we allow migration of already scheduled jobs to other machines upon the arrival of a new job, thus bridging the gap between online and offline setting. Already for a small amount of migration, our result compares with the best results to date in both online and offline settings. From the point of view of sensitivity analysis, our results imply that, only small changes are to be made to the current schedule to accommodate a new job, if we are satisfied with near optimal solution. The other online problem that we study is the well-known metrical task systems problem. We present a probabilistic analysis of the well-known text book algorithm called the work function algorithm. Besides average-case analysis we also present smoothed analysis, which is a notion introduced recently as a hybrid between worst-case and average-case analysis. Our analysis reveals that the performance of this algorithm is much better than worst-case for a large class of inputs. This motivates us to support smoothed analysis as an alternative model for evaluating the performance of online algorithms. The third problem that we investigate is a pursuit-evasion game: an algorithm (the pursuer) has to find/catch an adversary that is ‘hiding’ in a graph where both players can travel in the graph. This problem belongs to the rich field of search games and it addresses the question of how long it takes for the pursuer to find the evader in a given graph that, for example, corresponds to a computer network or a geographic terrain. Such game models are also used to design efficient communication protocols. We present improved results against adversaries with varying power and also present tight lower bounds.