CoRe: Constrained Reinforcement Learning for Network Control
- Reference number
- ID20-0027
- Project leader
- Tumova, Jana
- Start and end dates
- 201201-251201
- Amount granted
- 2 500 000 SEK
- Administrative organization
- KTH - Royal Institute of Technology
- Research area
- Information, Communication and Systems Technology
Summary
Automation in mobile networks has been gaining significant attention. So far, the monitoring and controlling of networks have been mainly conducted by domain experts and engineers. This project aims at new methods for autonomous decisions, e.g. base stations regarding power control and link adaptation in a mission critical context. The main focus will be on design and evaluation of decision-making mechanisms for future mobile networks that need to meet different Quality of Service (QoS) requirements. In recent years, Reinforcement Learning (RL) became a promising solution for dealing with optimal decision and control of agents that interact with uncertain environments. The research challenges that will be addressed are: i) the avoidance of potential actions that trigger locally sub-optimal decisions and lead to undesired states, where the network temporarily does not meet expected QoS (constrained exploration). ii) the synchronization of multi-agents' objectives since conflicting actions have to be handled in the context of QoS optimization (conflict resolution). iii) real-time distributed decision-making delays since agents in the network suffer from communication delays (handling delays). The results will improve the performance of mobile networks and reduce the operating cost based on the increased level of automation. The demonstrations will take place both in network simulators and in 5G testbed metropolitan networks that can be provided by Ericsson.
Popular science description
Over the last decades the demand of mobile networks that offer better and more stable quality of service and has been increasing. The reason of that is a higher usage of network devices (mobile phones, computers, etc.) and the recent boost of digitalization (more data should be transferred). Such demand has been observed both in Swedish society and industry as well where users desire a better quality of service. In order for the mobile networks to handle such demand, they need to be able to operate with high degree of automation. Network automation will allow the connectivity demand of the users and industries to be satisfied, a result that will have a great impact to our lives in Sweden and abroad. Traditionally, typical network operations such as monitoring, maintenance, management and control has been mainly conducted by field experts. This is no longer efficient to address the current demand by the users. In this context, this project aims to provide a framework that allow the networks to perform such operations autonomously, with minimum human intervention. This will be achieved through this project by a novel mathematical framework that allows the networks to perform decision making by their own. The proposed algorithms will be tested both in simulators and in live 5G testbed networks in order to make sure that they guarantee the desire quality of service by the users. According to the project plan, the results of this project will be available in live networks by the end of the project, i.e., by December 2025. The expected users of the outcome of this project will be all potential users of the networks what will operate autonomously under the proposed algorithms. In general, network users can be public people, companies, hospitals, connected vehicles etc. We strongly believe that this project will improve the current generation of communication networks and will boost the telecommunication market, which is one of the main sources of the economic growth of Sweden. As a result, it will have a huge impact in Swedish society.