I am proud to announce that the corrected version of my thesis “Bandit algorithms for searching large spaces" has been approved by my two examiners, Dr David Silver and Prof. Rémi Munos. Abstract below.
Bandit games consist of single-state environments in which an agent must sequentially choose actions to take, for which rewards are given. The objective being to maximise the cumulated reward, the agent naturally seeks to build a model of the relationship between actions and rewards. The agent must both choose uncertain actions in order to improve its model (exploration), and actions that are believed to yield high rewards according to the model (exploitation). The choice of an action to take is called a play of an arm of the bandit, and the total number of plays may or may not be known in advance.
Algorithms designed to handle the exploration-exploitation dilemma were initially motivated by problems with rather small numbers of actions. But the ideas they were based on have been extended to cases where the number of actions to choose from is much larger than the maximum possible number of plays. Several problems fall into this setting, such as information retrieval with relevance feedback, where the system must learn what a user is looking for while serving relevant documents often enough, but also global optimisation, where the search for an optimum is done by selecting where to acquire potentially expensive samples of a target function. All have in common the search of large spaces.
In this thesis, we focus on an algorithm based on the Gaussian Processes probabilistic model, often used in Bayesian optimisation, and the Upper Confidence Bound action-selection heuristic that is popular in bandit algorithms. In addition to demonstrating the advantages of the GP-UCB algorithm on an image retrieval problem, we show how it can be adapted in order to search tree-structured spaces. We provide an efficient implementation, theoretical guarantees on the algorithm’s performance, and empirical evidence that it handles large branching factors better than previous bandit-based algorithms, on synthetic trees.