Combinatorial Bandits for Sequential Learning in Colonel Blotto Games

The Colonel Blotto game is a renowned resource allocation problem with a long-standing literature in game theory (almost 100 years). In this work, we propose and study a regret-minimization model where a learner repeatedly plays the Colonel Blotto game against several adversaries. At each stage, the learner distributes her budget of resources on a fixed number of battlefields to maximize the aggregate value of battlefields she wins; each battlefield being won if there is no adversary that has higher allocation. We focus on the bandit feedback setting. We first observe that it can be modeled as a path planning problem. It is then possible to use the classical ComBand algorithm to guarantee a sub-linear regret in terms of time horizon, but this entails two fundamental challenges: (i) the computation is inefficient due to the huge size of the action set, and (ii) the standard exploration distribution leads to a loose guarantee in practice. To address the first, we construct a modified algorithm that can be efficiently implemented by applying a dynamic programming technique called weight pushing; for the second, we propose methods optimizing the exploration distribution to improve the regret bound.