This thesis examines the design of cognitive routing to improve wireless ad hoc network performance in terms of throughput and delay, as well as reducing the impact of relaying on the network, without deteriorating end-to-end capacity. Routing metrics are designed to replace the conventional shortest path routing metric (hop count) used in many existing routing protocols (e.g. AODV, DSR and DSDV). The routing metrics take into account a node/link’s surrounding environment conditions (such as disturbed node number, bottleneck capacity, and channel utilization).
A new family of Disturbance/Inconvenience based Routing (DIR) metrics are proposed initially, which takes both inward and outward interference into account in their weight metric designs, in order to reduce the impact of interference in the crowded areas. Then, a Bottleneck-Aware Routing (BAR) metric is developed to reduce bottleneck node problems for wireless ad hoc networks. BAR not only takes an individual node’s interference and capacity into account but is also aware of the location of bottleneck nodes such that routes can be intelligently established to avoid congested areas, and especially avoid bottlenecks. Under low traffic load conditions, both metrics show a significant reduction in the congestion levels compared with the shortest path routing metric despite increasing the relaying burden on nodes in the network.
Taking these findings into account, a cross-layer design is developed, where a Cognitive Greedy-Backhaul (CGB) routing metric is combined with a Reinforcement Learning based Channel Assignment Scheme (RLCAS). By applying a reinforcement learning algorithm to the channel assignment scheme, channels can be assigned through a more distributed and efficient approach. Moreover, the hidden node problem is mitigated and better channel spatial reuse is achieved due to the learning within the channel assignment scheme. By obtaining cross-layer information from the channel assignment scheme, CGB incorporates channel utilization into its metric in order to build backhaul links, while still utilising relatively short paths to help reduce the relaying burden. Thus, two significant advantages can be achieved using this cross-layer design: limiting the relaying impact and maintaining network capacity for wireless ad hoc networks. Results show that this cross-layer design outperforms the other schemes in terms of energy consumption, throughput and delay under varying traffic loads. Furthermore, the learning progress of RLCAS is studied in a multi-hop scenario and the reason why the learning engine of RLCAS performs well when it is associated with the CGB routing metric is also provided.
