Selected Topic: Artificial Intelligence
Research groups that are working with 5G Projects
Connected devices have wildly different traffic patterns, and with 5G these patterns are expected to become even more different due to better infrastructure allowing support of different latencies, throughput and other Quality of Service (QoS) requirements. In such a scenario, the already crowded spectrum will end up excessively populated as there will be far more devices competing for transmission spectrum, on those grounds resource allocation needs to guarantee a minimum level of service for users in crowded places.
Resource scheduling is a complicated problem due to multiple goal optimization that are mutually exclusive, including, but not limited to, latency, throughput, fairness and spectrum efficiency . A lot of heuristics are already in use to solve and improve block allocation, one of the simplest and most naive solution is Round-Robin scheduler, and a bit more complex ones with Genetic Algorithms (GA)  or other techniques.
Optimal solution would require knowledge of the future to allow scheduling of resources in optimal manner, that is impossible to be done in real-time, and even if would be possible due to the sheer amount of data there would need to be a tremendous amount of processing power to solve the problem in the short amount of time imposed by mobile networks standards. One possible solution comes by utilizing Artificial Neural Networks (ANN) to find a sub optimal solution training with real traffic data.
The 5G-Range Project expects 5G networks to reach 50Km range of coverage with focus in rural areas, where traffic is far different in comparison to urban areas. Moreover in rural areas there is an abundance of white-spaces in the spectrum, that is caused by lower demand over the licensed channels as well as in the unlicensed bands, making a huge improvement in spectrum efficiency if those holes can be used for something positive, such as Device to Device (D2D) or Machine to Machine (M2M) communications, it can also be used for traffic offloading from mobile network to unlicensed band trough Wi-Fi or any other wireless technology from that band. On these facts a traffic generator, which is based on QoS Class Identifier (QCI) values, that range from 0 to 15 and can be used to classify the type of application .
For mobile networks, an ANN with unsupervised learning can be used with reinforcement learning, where a fitness value or reward is given to each action, prediction, done by the ANN, by doing that operation several times and going through the data. In different epochs, the neural network is able to learn fairly well how it should allocated resource blocks based on how much reward it is getting as feedback . Manipulating the reward system more weight can be given for a specific characteristic that one might need to optimize, such as spectrum efficiency or latency. Current reward system is as follows:
This system has perfect value if all characteristics are 1, anything out, i.e. a miss, will make the reward go down in the direction where it is failing, another good thing is that weights can be put in the actual Reward equation to give priority to some attributes.
The current state of this work is at verifying different traffic profiles that should provide datasets of good value for the context of 5G-Range project. A study conducted in South Africa has shown traffic characterization in that country, this work analyzed traffic for 10 days and has found that 68.45% is Web Traffic and the next big portion is Video, such as Youtube (Buffered) or VoIP systems (livestreams) . Based on that study a dataset is being created covering the percentages of internet usage and also the user per kilometer requirement of the 5G Range Project of 2 users/km.
 Okvist, Simonsson and Asplund. LTE Frequency Selective Scheduling Performance and Improvements Assessed by Measurements. International Symposium on Personal, Indoor and Mobile Radio Communications, 2011.
 Yang, Xu, Han, Rehman and Tao. GA Based Optimal Resource Allocation for Device to Device. WCNC 2014 - Workshop on D2D and Public Safety Communications, 2014.
 3GPP. Policy and Charging Control Architecture. TS 23.203, 2018.
 Arulkumaran, Deisenroth, Brundage and Bharath. A Brief Survey of Deep Reinforcement Learning. IEEE Signal Processing Magazine Special Issue on Deep Learning for Image Understanding, 2017.
 Johnson, Pejovic, Belding and Stam. Traffic Characterization and Internet Usage in Rural Africa. WWW '11 Conference companion on World Wide Web, 2011.