RRP: Reinforced Routing Policy Architecture for MANET Routing
Aviel Glam, Barak Farbman, Ariel Shleifer · 2019 IEEE International Conference on Microwaves, Antennas, Communications and Electronic Systems (COMCAS) · 2019
In this paper we present a novel architecture for obtaining a routing scheme for mobile ad-hoc networks (MANETs). The goal as always in routing is obtaining a well-suited routing table for maximization of the message completion ratio (MCR) while keeping the delay low. We have designed a system architecture that uses an extended routing table and any side information available to the node. Side information can come in the form of physical layer indications (SNR, RSSI, etc.) or congestion (e.g. output queues length) for example. We modeled a highly dynamic MANET environment as a Markov Decision Process (MDP) and deployed reinforcement learning (RL) techniques from training database augmentation, hybrid reward, adaptive learning rate and assisted learning methodology (setting bounds on the learning process and outcome). The resulting routing policy has a probabilistic element regarding transmission to every next-hop option and it out-performs current policies both in different network loads and different noise environments i.e. the learned policy has higher MCR and almost the same delay as previously known algorithms.