Experimental validation of a reinforcement learning based approach for a service-wise optimisation of heterogeneous wireless sensor networks

Due to their constrained nature, wireless sensor networks (WSNs) are often optimised for a specific application domain, for example by designing a custom medium access control protocol. However, when several WSNs are located in close proximity to one another, the performance of the individual networks can be negatively affected as a result of unexpected protocol interactions. The performance impact of this `protocol interference' depends on the exact set of protocols and (network) services used. This paper therefore proposes an optimisation approach that uses self-learning techniques to automatically learn the optimal combination of services and/or protocols in each individual network. We introduce tools capable of discovering this optimal set of services and protocols for any given set of co-located heterogeneous sensor networks. These tools eliminate the need for manual reconfiguration while only requiring minimal a priori knowledge about the network. A continuous re-evaluation of the decision process provides resilience to volatile networking conditions in case of highly dynamic environments. The methodology is experimentally evaluated in a large scale testbed using both single- and multihop scenarios, showing a clear decrease in end-to-end delay and an increase in reliability of almost 25 %.