Learning for adaptive anycast in vehicular delay tolerant networks

Abstract

We propose a protocol for vehicular delay tolerant networks (DTN) with special focus on anycast communications from a vehicle to the cloud where multiple gateways (such as road side units) exist. The protocol employs a Q-learning algorithm to estimate the multi-hop destination encounter probability by discounting the reward with the number of forwards (number of hops). Anycast encounter probability is maintained in a specific entry which considers multiple gateways as the same virtual destination. The proposed protocol also uses an adaptive data replication scheme to take into account the destination encounter probability and the relative velocity between vehicles jointly, which can achieve a high data delivery ratio and low overhead. We use computer simulations to evaluate the proposed protocol.