Distributed Disaggregated Communications via Reinforcement Learning and Backpressure (D2CRaB)

Abstract

This is a practitioner paper focused on the real problem of effectively communicating within distributed and disaggregated multi-domain battlespace operational environments. To do so, communications systems must be adaptive in response to ever-changing on-mission events such as, network congestion (e.g., degraded comms links and excessive network traffic), enemy interference (e.g., jamming, cyber-attacks), line of Sight degradation (e.g., weather conditions). These challenges cause networks to drop packets, and presently they do so according to pre-specified Quality of Service (QoS) policies. However, these QoS policies alone do not actually mitigate the congestion problem because if the impacted streams employ the reliable communication protocols, such as TCP, packets being dropped will be retransmitted over the same route. Thus, this approach can make the congestion problem even worse and waste the valuable bandwidth.The approach presented here introduces the novel Distributed Disaggregated Communications via Reinforcement Learning (RL) and Backpressure (D2CRaB) scheme to address the above stated problem. D2CRaB accomplishes this in two ways: (1) by bridging route selection and congestion control via the backpressure scheme and (2) by leveraging RL to enable the dynamic and continuous adaptation of network situation.