AFEDA: Enhancing Network Slices Acceptance Ratio with Transformer-based Feature Extraction

The advent of 5G network slicing technology makes it possible to divide a shared infrastructure into logical networks called network slices, each providing a customized quality of service (QoS). Consequently, the QoS satisfaction has shifted from "how to provide tailored QoS?" to "how to improve the acceptance ratio of deployed network slices?". Existing works have successfully employed Deep Reinforcement Learning (DRL) agents to address this challenge, proposing various placement strategies guided by reward functions. In this paper, we introduce AFEDA, an improved DRL agent by coupling it with a transformer-based active features extractor and demonstrate that beyond the reward function, extracting active features from observations significantly helps a DRL agent to enhance the acceptance ratio of slices. Through extensive simulations on two infrastructures, we show that AFEDA is capable of placing 14% to 31% more slices compared against a DRL agent with same configurations while using the same observations as raw features.