Replicative Simulator: A Step Towards Harmonizing AI and Wireless Communication Systems
An innovative way to apply artificial intelligence technology in 5G wireless environments
Advanced Comm. Research Center,
Advanced Comm. Research Center,
Advanced Comm. Research Center,
The evolutions of cellular technologies over multiple generations have increased the complexity of the systems. In particular, the recently commercialized 5G is vastly more complicated compared to LTE in terms of the range of spectrum, the number of physical layer parameters, and the complexity of base station and network architecture in order to fulfill enhanced requirements on all aspects of the system as in Figure 1. Manually tuning such systems for better user experience or lower operational costs becomes infeasible due to the complexity; hence it highlights the needs for optimal, adaptive, and scalable management of the system that are automatic. For this reason, automatic optimization of cellular systems by applying Artificial Intelligence (AI) or Machine Learning (ML) is being discussed as a promising approach for 5G and upcoming 6G systems.
AI/ML for Mobile Networks
The recently emerged AI/ML technologies have shown great successes in dealing with complexities in multiple areas such as computer vision and voice recognition. These technologies are suitable for recognizing highly complex and non-linear relations hidden within data and leveraging them to carry out required tasks. Naturally, the adoption of AI/ML is expanding to many different areas with complexities such as autonomous driving, robotics, smart factory, and finance, which in turn is expediting the evolutions of corresponding systems and services.
Among many AI/ML technologies, Reinforcement Learning (RL) in particular is considered as a promising approach for many problems in managing complex cellular systems. RL aims to find optimal controls of complex systems in changing conditions by learning from the interactions between the systems and their environments. It can handle highly complex situations such as cellular systems and their interworking with users under different conditions, but RL also requires a large amount of various trial-and-error experiences for it to learn. Therefore, a cost-efficient way to produce a large amount of experience or data is a prerequisite for leveraging RL.
Reinforcement Learning(RL) for Networks
An RL agent can interact with real systems to train itself, but doing so can cause instability of the system and disastrous user experiences when the RL agent takes unexplored routes to learn about the system, which is called ‘exploration’ and is essential process for successful RL model training. In addition, the experience from a single physical system can be limited and biased for an agent to learn general trends. To overcome such limitations, simulators are widely used for RL training as they can offer a large amount of experience without the danger of compromising real systems serving users. However, simulation also has its own limitations. The main drawback of the simulation comes from the fact that it is not the real system, and it may not sufficiently reflect the complex interactions of the real world systems and environments. Therefore, closing the gap between simulations and real-world phenomena is an indispensable task to be able to train and use AI/ML for this type of complex systems.
Digital Twin for Networks
The use of simulators for optimizing real-world system is not a new approach. The idea is widely used in many different industries and one of the popular concepts, called ‘Digital Twin,’ for factory optimization shown in Figure 2. It involves modeling real world to construct virtual world, running simulations in virtual world to find optimal controls, and then applying the controls back to the real-world system. This process can iterate until the predefined requirements are satisfied.
Figure 2: Digital twin conceptual architecture.
Ref : https://www2.deloitte.com/us/en/insights/
The same concept can be applied to optimize cellular systems as well. It means bridging the real world cellular systems and virtual world simulations, experimenting in the virtual world (simulations) to find optimal controls, and applying the findings back to the real world (real cellular systems). Creating this iterative loop would make possible to construct RL models for controlling complex cellular systems. The challenges include finding the right level of abstraction for the simulation, replicating the states of real world in virtual world, and modeling highly varying and non-deterministic real world responses in virtual world.
Samsung has initiated research on replicative simulators for cellular systems. A replicative simulator takes real-world observations, and reproduces system states and responses in simulation that match the observations. This reproducing capability enables us to not only apply AI/ML to cellular systems but also analyze what-if scenarios, validate new algorithms, etc., which offers a virtual testbed of cellular systems.
Figure 3: Replicative simulation bridging between the physical and virtual world.
The replicative simulators can be used to address traditional network optimization problems such as hand-over, load balancing, and energy saving. The optimization is typically performed to achieve goals such as improving user experience or reducing operational costs while avoiding degradation of Key Performance Indicators (KPIs). It can be done using heuristics, algorithms, or RL models, but we found that the RL outperforms other approaches when the complexities of system and requirement are high. The replicative simulation also offers validation opportunities that are typically not available in conventional setups. For example, it can be used to reconstruct a bad incident observed in real system and to validate if a specific modification of the system would have avoided the occurrence of the incident. Simulations as virtual testbeds can accelerate the speed of development cycle for they can run in parallel and in large-scale.
In Samsung, we have developed replicative simulators for cellular systems and are making progresses to leverage them in a way that would accelerate AI/ML adoption, revolutionize development process, and re-invent live site management.
About Samsung Research
Samsung Research is the advanced research and development (R&D) hub of Samsung Electronics’ *SET (End-products) Business which includes Consumer Electronics (CE) Division and IT & Mobile Communications (IM) Division. The hub leads the development of the future technologies with talented researchers and developers working in global R&D centers. For more information, please visit Samsung Research at https://research.samsung.com.