Information Bottleneck-Based Domain Adaptation for Hybrid Deep Learning in Scalable Network Slicing

Network slicing empowers operators to economically accommodate a diverse array of applications on a shared physical infrastructure. The escalating network deployments raise high system complexity, particularly in the presence of nontrivial inter-cell interference. This complexity necessitates more efficient methods to dynamically optimize resource management. Numerous solutions make use of deep learning techniques to deal with these challenges. However, deep learning models are constrained by their limited capacity for generalization and
adaptability when confronted with dynamic slicing configurations. In this paper, we come up with a novel hybrid learning approach IDLA to address this limitation and improve the scalability of slicing resource partitions by integrating deep learning, which provides high flexibility, and a non-linear optimization method, which exhibits high robustness across scenarios. Moreover, we proposed a DA approach under the context of TL to further extend the generality of IDLA with the help of IB theory. This DA method utilizes a VIB model to capture invariant patterns among disparate network environments, which are characterized by variant slice configurations. Specifically, this DA empowered IDLA algorithm for slicing resource partitioning consists of three-steps approach: 1) First, based on the samples collected under the revised slicing configurations (marked as source domains in TL), we pre-train a VIB-based slice QoS estimation model; 2) The derived VIB model is then deployed to the expected slicing scenario as target domain to resolve slice resource partitions with the IDLA, and collects network performance samples for model finetuning; 3) Lastly, the VIB estimator is finetuned on the mixture
of samples from source and scarce target samples, and step 2) is repeated until convergence. To assess the generality and flexibility of proposed solutions, we implemented the proposed IDLA algorithm with VIB-based model under a multi-cell dynamic slicing network scenario where the number and type of slices are changable under operation. In comparison with state-of-the-art solutions, including a DRL approach, our VIB-aided IDLA solution exihibits 10× faster convergence and 18% higher asymptote after slicing configuration changes. For transferability evaluation, we comprehensively investigated the VIB model’s accuracy under 4 pairs of slicing settings with various discrepancies in comparison with a simple DNN model and the one embedded with a sample reweighting technique for sample imbalance. The numerical results demonstrate that the VIB provides the highest estimation accuracy on target domains with different mix ratios of target domain samples, while maintains robust source domain performance.