Radio Resource Management by Augmenting Cellular Data
with Non-Cellular Attributes
Mobile device adoption is continually and exponentially growing while cellular providers have access to fragmented and limited amounts of spectrum available to them. Prediction of traffic in cellular networks is the first of the promising methods proposed to improve resource utilization and allocation among Base Stations. With the increasing number of publicly available environmental information, we can understand the surrounding environment at a finer granularity and utilize this open-source information to first predict the traffic at and second assign channels to Base Stations at given times to achieve a higher QoS.
Predicting the traffic at a given location helps us understand what number of resources are needed at a specific location at a given time and would otherwise be underutilized elsewhere. Opensource APIs allow exploring the environment from different angles, such as transceiver locations, busy periods for local establishments, and how many civilians live in a particular location.
The first part of the project for our Dynamic Channel Allocation (DCA) problem is we need to get accurate cellular traffic predictions utilizing a Long- Short-Term Memory (LSTM) model that is trained with historical Call Detail Records (CDRs) provided by Telecom Italia (TIM) and popular location data from the Google Places API. We have collected Base Station transceiver location estimations from OpenCellID’s opensource API and, using a vanilla K-Means clustering algorithm with 100 iterations on the transceivers’ latitude and longitude, we obtained estimated Base Station locations (lat/lng pairs). To gauge how many user equipment (UEs) are in each location, the population for Milan was obtained through Facebook’s public API.
The second part of the project for our DCA problem is to establish a simulation environment based on the open-source information collected in the first part. Through the use of Network Simulator 3’s (NS3) simulation environment, we are able to represent an LTE network close to 3GPP’s standards and obtain quality service metrics at the physical layer. To attach deep learning to our network, we are using the NS3-GYM attachment to implement a Deep Q-Network to determine which channels should be used at each Base Station at a given time, based on past traffic and QoS reports. The NS3-GYM and NS3 environments are hosted on an image on the Orbit Testbed with the codebase stored in a Version Control System (VCS) for future reproducibility.
Accurate traffic predictions have been achieved through the trained LSTM model on both cellular and noncellular attributes and published in . The current focus is utilizing these noncellular attributes with traffic predictions to optimize resource allocation, including dynamic channel allocation.
Kuber, T., Seskar, I., & Mandayam, N. (2021, March). Traffic prediction by augmenting cellular data with non-cellular attributes. In 2021 IEEE Wireless Communications and Networking Conference (WCNC) (pp. 1-6). IEEE.