Online Combinatorial Optimization under Bandit Feedback

Sammanfattning: Multi-Armed Bandits (MAB) constitute the most fundamental model for sequential decision making problems with an exploration vs. exploitation trade-off. In such problems, the decision maker selects an arm in each round and observes a realization of the corresponding unknown reward distribution. Each decision is based on past decisions and observed rewards. The objective is to maximize the expected cumulative reward over some time horizon by balancing exploitation (arms with higher observed rewards should be selectedoften) and exploration (all arms should be explored to learn their average rewards). Equivalently, the performanceof a decision rule or algorithm can be measured through its expected regret, defined as the gap betweenthe expected reward achieved by the algorithm and that achieved by an oracle algorithm always selecting the bestarm. This thesis investigates stochastic and adversarial combinatorial MAB problems, where each arm is a collection of several basic actions taken from a set of $d$ elements, in a way that the set of arms has a certain combinatorial structure. Examples of such sets include the set of fixed-size subsets, matchings, spanning trees, paths, etc. These problems are specific forms of online linear optimization, where the decision space is a subset of $d$-dimensional hypercube.Due to the combinatorial nature, the number of arms generically grows exponentially with $d$. Hence, treating arms as independent and applying classical sequential arm selection policies would yield a prohibitive regret. It may then be crucial to exploit the combinatorial structure of the problem to design efficient arm selection algorithms.As the first contribution of this thesis, in Chapter 3 we investigate combinatorial MABs in the stochastic setting and with Bernoulli rewards. We derive asymptotic (i.e., when the time horizon grows large) lower bounds on the regret of any algorithm under bandit and semi-bandit feedback. The proposed lower bounds are problem-specific and tight in the sense that there exists an algorithm that achieves these regret bounds. Our derivation leverages some theoretical results in adaptive control of Markov chains. Under semi-bandit feedback, we further discuss the scaling of the proposed lower bound with the dimension of the underlying combinatorial structure. For the case of semi-bandit feedback, we propose ESCB, an algorithm that efficiently exploits the structure of the problem and provide a finite-time analysis of its regret. ESCB has better performance guarantees than existing algorithms, and significantly outperforms these algorithms in practice. In the fourth chapter, we consider stochastic combinatorial MAB problems where the underlying combinatorial structure is a matroid. Specializing the results of Chapter 3 to matroids, we provide explicit regret lower bounds for this class of problems. For the case of semi-bandit feedback, we propose KL-OSM, a computationally efficient greedy-based algorithm that exploits the matroid structure. Through a finite-time analysis, we prove that the regret upper bound of KL-OSM matches the proposed lower bound, thus making it the first asymptotically optimal algorithm for this class of problems. Numerical experiments validate that KL-OSM outperforms state-of-the-art algorithms in practice, as well.In the fifth chapter, we investigate the online shortest-path routing problem which is an instance of combinatorial MABs with geometric rewards. We consider and compare three different types of online routing policies, depending (i) on where routing decisions are taken (at the source or at each node), and (ii) on the received feedback (semi-bandit or bandit). For each case, we derive the asymptotic regret lower bound. These bounds help us to understand the performance improvements we can expect when (i) taking routing decisions at each hop rather than at the source only, and (ii) observing per-link delays rather than end-to-end path delays. In particular, we show that (i) is of no use while (ii) can have a spectacular impact.For source routing under semi-bandit feedback, we then propose two algorithms with a trade-off betweencomputational complexity and performance. The regret upper bounds of these algorithms improve over those ofthe existing algorithms, and they significantly outperform state-of-the-art algorithms in numerical experiments. Finally, we discuss combinatorial MABs in the adversarial setting and under bandit feedback. We concentrate on the case where arms have the same number of basic actions but are otherwise arbitrary. We propose CombEXP, an algorithm that has the same regret scaling as state-of-the-art algorithms. Furthermore, we show that CombEXP admits lower computational complexity for some combinatorial problems.