Design of AI-based Resource Forecasting Methods for Network Slicing

Abstract

With the forthcoming of 5G networks, the underlying infrastructure needs to support a higher number of heterogeneous services with different QoS needs than ever. For that reason, 5G inherently provides a way to allocate these services over the same infrastructure through the concept of Network Slicing. However, to maximize revenue and reduce operational costs, a method to proactively adapt the resources assigned to each slice becomes imperative. For that reason, this work presents two Machine Learning (ML) models, leveraging Long-Short Term Memory (LSTM) and Random Forest algorithms, to forecast the throughput of each slice and adapt accordingly the amount of resources needed. The models are evaluated using NS-3, which has been integrated with the ML models through a shared memory framework. This enables a closed loop in which the predictions of the models can be used at run time to introduce changes in the network. Consequently, it makes it able to cope with the forecasted requirements, eliminating the need for off-line training and resembling better a real-life scenario. The evaluation performed shows the ability of the models to predict the slices' throughput under various settings and proves that Random Forest provides up to 26% better results than LSTM.

ECP: Error-Aware, Cost-Effective and Proactive Network Slicing Framework

“…Like medical load forecasting, predictive deep learning models (e.g., LSTM) have outperformed statistical models in network load prediction in different scenarios [19], [20]. A recent study by [21] used LSTM and Random Forest to predict network slicing KPIs compliance but did not address automated resource fine-tuning for varying resource needs, which we address. Additionally, many of the literature works, e.g., [22], [23] focused only on optimizing the resources without the inclusion of the routing in their optimization, which we have incorporated.…”

“…Finally, if the agent finds the optimal configuration per all slices that minimizes the total cost and abides by all constraints for the actual load, it will be rewarded, as seen in the last row of r t . Thus, (21) represents the reward function:…”

“…After ECP initialization, the training phase starts, and it's categorized into data collection (lines 4-20) and PPO's learning stages (lines [21][22][23][24][25]. Data collection starts by iterating over a specific number of training hours per episode (e.g., 12 hours from 8 AM to 8 PM); these hours represent a timestep where there is a load for a set of services S which we need to optimize.…”

“…The second test (lines 13-15) checks if the summation of allocated resources across each node is higher than its limit; if so, then the resources of the competing services will be divided proportionally among the demanding services (i.e., More resources are given to the service with higher demand) and M is updated. After a t is checked and M is updated, the agent performs the allocation, the reward is calculated as in (21), and the trajectory is saved in the buffer B.…”

“…After the data collection finishes, the PPO learning stage starts (lines [21][22][23][24][25], where a mini-batch m is sampled from the buffer, and three main steps follow. The first is computing the reward-to-Go Rt per each trajectory in m. Next is computing the advantage estimate Âπ k t that identifies how an action is better than others in a given state and accordingly updates PPO main policy (actor's NN) and value-function (critique NN) networks.…”