Abstract: Modern network systems give rise to many challenging control and resource management problems, from congestion control to video bitrate adaptation to scheduling jobs in computer clusters. Classical approaches to these problems, developed over the last four decades, rely on algorithms designed by human experts. But as network systems have grown in complexity, it has become exceedingly difficult to design control algorithms that perform well in a variety of conditions. In this talk I will discuss some of our work on systems that use reinforcement learning (RL) to learn highly-optimized control algorithms entirely through experience. I will illustrate this approach in two systems, Pensieve and Decima, that use RL to learn, respectively, network-specific bitrate adaptation policies for video streaming and workload-specific scheduling policies for data processing clusters. Finally, I will highlight some of the broader challenges for applying RL to network systems, drawing on lessons learned in building systems like Pensieve and Decima. I will discuss how these challenges can in some cases motivate new learning techniques, using our recent work on RL algorithms for “input-driven” environments as an example.
Bio: Mohammad Alizadeh is an Associate Professor of Computer Science at MIT. His research interests are in the areas of computer networks and systems, and applied machine learning. His current research focuses on learning-augmented systems, video streaming, and congestion control algorithms for datacenter and wide-area networks. Mohammad’s research has garnered significant industry interest. His work on datacenter transport protocols has been implemented in Linux and Windows, and has been deployed by large network operators; his work on adaptive network load balancing algorithms has been implemented in Cisco’s flagship datacenter switching products. Mohammad received his Ph.D. from Stanford University and then spent two years at Insieme Networks (a datacenter networking startup) and Cisco before joining MIT. He is a recipient of the Microsoft Research Faculty Fellowship, VMware Systems Research Award, NSF CAREER Award, SIGCOMM Rising Star Award, Alfred P. Sloan Research Fellowship, and multiple best paper awards.