Branislav Kveton, Jia Yuan Yu, Georgios Theocharous, Shie Mannor
In this paper, we study a sequential decision making problem. The objective is to maximize the average reward accumulated over time subject to temporal cost constraints. The novelty of our setup is that the rewards and constraints are controlled by an adverse opponent. To solve our problem in a practical way, we propose an expert algorithm that guarantees both a vanishing regret and a sublinear number of violated constraints. The quality of this solution is demonstrated on a real-world power management problem. Our results support the hypothesis that online learning with convex cost constraints can be performed successfully in practice.
Subjects: 12. Machine Learning and Discovery; 16. Real-Time Systems
Submitted: Apr 15, 2008