The unprecedented surge in data traffic experienced over the last decade has stretched the telecommunications networks to their capacity. According to Cisco’s forecast (included in the Cisco Visual Networking Index: Forecast and Methodology report), global IP traffic will have increased 127-fold from 2005 to 2021. This demand rise will even be exacerbated for wireless and mobile traffic, which is expected to account for more than 60% of the total IP traffic by 2021, growing twice as fast as fixed IP traffic.
However, the tremendous strain put on the mobile communications industry to meet new challenging technical requirements (in terms massive connectivity, broadband communications or ultra-reliability and low latency) cannot be decoupled from the increasing costs incurred by mobile network operators. On the one hand, the requirements imposed by the wide range of use cases and services have driven the standardization of a new radio interface (5G NR), the densification of the network, the allocation of new spectrum bands, etc. On the other hand, the need to reduce Operational and Capital Expenditure (OPEX and CAPEX) has laid the basis for network sharing and multitenancy in the framework of 5G networks.
When comparing 5G networks with previous mobile networks, there is a move from simple hierarchical networks to extremely complex and heterogeneous networks, shared by multiple operators and giving service to applications with completely different throughput, delay and/or jitter constraints. This complexity, which is inherent to 5G, makes essential to design new network management techniques able to cope with the 5G reality. Is this the right time for Machine Learning?
Machine Learning (ML) is defined as a subset of concepts and methods of the Artificial Intelligence (AI) field of knowledge that relies on the ability of computer systems to learn from existing data (input) and make decisions (output) to optimize an objective function. ML has gained momentum thanks to the explosion of data availability and the increasing complexity of computer systems. Although ML has not yet been applied extensively in communications networks, and particularly in mobile networks, it is the pillar of other scientific fields such as the computer vision field. Machine Learning can be divided into three main groups: Supervised Learning, Unsupervised Learning and Reinforcement Learning.
Supervised Learning (SL): As all ML methods in general, SL is aimed to learn a set of features from a data set. In this case, however, the participation of a supervisor is needed. The information required by SL algorithms is composed of a set of inputs and a set of outputs. SL, after a training phase, is able to tailor a predictive model with which any output can be estimated from the given input. SL is useful in systems where both inputs and outputs are known, but the relationship between them is unknown.
Unsupervised Learning (UL): The fundamental idea of UL is the analysis of unlabelled and uncategorized data to extract features without prior training. Given the characteristics of UL, it is able to perform more complex processing tasks than SL, since these techniques are mainly aimed to figure out unknown patterns for which no output is available (that is the reason why they are not supervised). Some applications are connected with system calibration, unusual result/faults detection, etc.
Reinforcement Learning (RL): RL relies on the creation of an agent which, based on the interactions with the environment, takes proper decisions to maximize a reward function. Contrary to SL, RL does not provide the optimal action to take, but a sequence of temporal decisions according to rewards/punishment. RL assumes a Markov decision process where an agent might ignore a reward in order to explore better decisions.
ML can be applied to a variety of open research challenges in the framework of 5G. The potential of ML in future complex and varying mobile networks has been already pointed out by the research community. However, the effectiveness of applying ML to mobile networks will be tightly coupled with the characteristics of the problem and the selected ML technique. In that sense, the matching between each 5G network problem and the most convenient ML technique is still an open issue.
Network reconfigurability at different levels is one of the most promising applications of ML to 5G networks. The prediction of temporal and geographical traffic dynamics will be a key aspect in the design of slicing algorithms, online caching policies or dynamic functional split. For instance, SL techniques will be appropriate to learn spatial and temporal patterns to proactively tailor virtual networks on top of the physical shared radio access networks. UL, for its part, has a significant potential to reveal statistical features, such as multimedia content popularity or users’ preferences. Finally, RL will play a key role in the short-term adaptation such as the radio resources management, dual connectivity management, etc.
As explained above, it is apparent and a common belief in the mobile communications community that ML will play a key role in future mobile networks. Its ability to learn from existing data and make decisions with minor human intervention will revolutionize the way networks are conceived. Thus, ML-assisted 5G networks will mean a technological breakthrough and SPOTLIGHT aims to shed light on it.