UMass Machine Learning and Friends Lunch | Main / Reliable Reinforcement Learning For Resource Management Problems

Algorithms that reliably combine machine learning with optimization can be applied to many complex domains, including resource management and planning under uncertainty, with little human effort. Developing such algorithms is a long-standing challenge in reinforcement learning and operations research. Current reinforcement learning algorithms usually approximate the solution iteratively based on a set of samples and features. Although these iterative algorithms can achieve impressive results in many domains, they have drawbacks that significantly limit their applicability; they often require extensive parameter tweaking to work well, and provide only weak guarantees on the solution quality. To alleviate these drawbacks, we propose and study new optimization-based approximate methods. Because these optimization-based methods decouple the solution properties from the algorithm, they are easy to analyze, offer much stronger guarantees than iterative algorithms, and have no parameters to tweak. We also show that the strong guarantees of optimization-based methods can be used to adaptively choose appropriate features, a crucial component in achieving flexible algorithms. Because the optimization-based algorithms also perform well on resource-management and benchmark problems, they are an increasingly-attractive alternative to traditional iterative methods.