Efficiently reducing buffer sizes in Internet routers may be one of the main outstanding open problems in networking. The networking community has worked on small buffers for nearly 20 years, even organizing dedicated workshops, albeit with limited impact on real routers.
Our goal is to use reinforcement learning to design active queue management policies that can enable small router buffers with good throughput, loss, and delay characteristics.
We want to learn the characteristics of different congestion control algorithms so as to be able to quickly recognize the congestion control used by any flow at any router.
Our goal is to provide a dynamic multihoming solution where we decide how to best connect to the Internet (ADSL, cable, LTE, etc.) to get to a website. We look at whether multi-armed bandits approaches can help solve this problem, although it seems that the non-stationary characteristics of the problem are hard to solve using AI approaches.
