Abstract: Cellular offloading in device-to-device communication is a challenging
optimisation problem in which the improved allocation of radio resources can
increase spectral efficiency, energy efficiency, throughout and reduce latency.
The academic community have explored different optimisation methods on these
problems and initial results are encouraging. However, there exists significant
friction in the lack of a simple, configurable, open-source framework for
cellular offload research. Prior research utilises a variety of network
simulators and system models, making it difficult to compare results. In this
paper we present GymD2D, a framework for experimentation with physical layer
resource allocation problems in device-to-device communication. GymD2D allows
users to simulate a variety of cellular offload scenarios and to extend its
behaviour to meet their research needs. GymD2D provides researchers an
evaluation platform to compare, share and build upon previous research. We
evaluated GymD2D with state-of-the-art deep reinforcement learning and
demonstrate these algorithms provide significant efficiency improvements.