Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal Online Learning of Safe Regions in Cloud Radio Access Networks

Kim Hammar, Tansu Alpcan, Emil Lupu

TL;DR

Causal Online Learning (COL) addresses safe online control in cloud RANs by identifying a safe operating region for dynamic resource configurations. It fuses causal inference to bootstrap an initial safe region with Gaussian-process-based Bayesian learning to progressively expand it through interventions, guided by an information-gain–per–cost acquisition rule. COL provides probabilistic safety guarantees during learning and convergence to the full safe region under standard assumptions, demonstrating up to 10x improvements in sample efficiency over non-causal baselines on a 5G testbed. The approach enables safe autonomous management of RANs and offers a path to extending safe online learning to other networked systems and digital twins.

Abstract

Cloud radio access networks (RANs) enable cost-effective management of mobile networks by dynamically scaling their capacity on demand. However, deploying adaptive controllers to implement such dynamic scaling in operational networks is challenging due to the risk of breaching service agreements and operational constraints. To mitigate this challenge, we present a novel method for learning the safe operating region of the RAN, i.e., the set of resource allocations and network configurations for which its specification is fulfilled. The method, which we call (C)ausal (O)nline (L)earning, operates in two online phases: an inference phase and an intervention phase. In the first phase, we passively observe the RAN to infer an initial safe region via causal inference and Gaussian process regression. In the second phase, we gradually expand this region through interventional Bayesian learning. We prove that COL ensures that the learned region is safe with a specified probability and that it converges to the full safe region under standard conditions. We experimentally validate COL on a 5G testbed. The results show that COL quickly learns the safe region while incurring low operational cost and being up to 10x more sample-efficient than current state-of-the-art methods for safe learning.

Causal Online Learning of Safe Regions in Cloud Radio Access Networks

TL;DR

Causal Online Learning (COL) addresses safe online control in cloud RANs by identifying a safe operating region for dynamic resource configurations. It fuses causal inference to bootstrap an initial safe region with Gaussian-process-based Bayesian learning to progressively expand it through interventions, guided by an information-gain–per–cost acquisition rule. COL provides probabilistic safety guarantees during learning and convergence to the full safe region under standard assumptions, demonstrating up to 10x improvements in sample efficiency over non-causal baselines on a 5G testbed. The approach enables safe autonomous management of RANs and offers a path to extending safe online learning to other networked systems and digital twins.

Abstract

Cloud radio access networks (RANs) enable cost-effective management of mobile networks by dynamically scaling their capacity on demand. However, deploying adaptive controllers to implement such dynamic scaling in operational networks is challenging due to the risk of breaching service agreements and operational constraints. To mitigate this challenge, we present a novel method for learning the safe operating region of the RAN, i.e., the set of resource allocations and network configurations for which its specification is fulfilled. The method, which we call (C)ausal (O)nline (L)earning, operates in two online phases: an inference phase and an intervention phase. In the first phase, we passively observe the RAN to infer an initial safe region via causal inference and Gaussian process regression. In the second phase, we gradually expand this region through interventional Bayesian learning. We prove that COL ensures that the learned region is safe with a specified probability and that it converges to the full safe region under standard conditions. We experimentally validate COL on a 5G testbed. The results show that COL quickly learns the safe region while incurring low operational cost and being up to 10x more sample-efficient than current state-of-the-art methods for safe learning.
Paper Structure (23 sections, 2 theorems, 28 equations, 14 figures, 2 tables, 1 algorithm)

This paper contains 23 sections, 2 theorems, 28 equations, 14 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Use Case: Identifying the Safe Region of a Cloud Radio Access Network
  4. System Model
  5. Problem: Online Learning of the Safe Region
  6. Our Method: Causal Online Learning (COL)
  7. Initializing the Safe Region through Causal Inference
  8. Bayesian Learning of the Safe Region
  9. Selecting Interventions through Active Learning
  10. Summary of Our Method for Identifying the Safe Region
  11. Illustrative Example
  12. Experimental Evaluation on a 5G Testbed
  13. Testbed Implementation
  14. Testbed Scenarios
  15. Experiment Setup
  16. ...and 8 more sections

Key Result

Proposition 1

Assume a) that samples in the dataset $\mathcal{D}_{t_k}$ are independent and identically distributed (i.i.d.); and b) that the difference between the causal effect $p_{\mathbf{U}'_{t_k}}(\mathbf{u}'_{t_k})$ and the (prior) mean function $m_{\mathbf{U}'_{t_k}}$ lies in the reproducing kernel Hilbert for all $\mathbf{u}'_{t_k} \in U'$, where $\hat{f}_{\mathbf{U}'_{t_k}} = E_{f \sim (\hat{f}_{\mathb

Figures (14)

  • Figure 1: Causal Online Learning (COL): a method for identifying the safe region of a cloud radio access network (RAN) where virtual baseband units (vBBUs) are distributed across an (edge) cloud infrastructure. The safe region contains the set of control inputs (e.g., resource allocations) for which the RAN's specification (i.e., service requirements and constraints) is fulfilled. COL exploits the causal structure of the RAN to learn the safe region from system observations, which are collected through a sequence of interventions.
  • Figure 2: Architecture of the target system that we study in this paper: a 5G cloud radio access network (RAN). The network follows the O-RAN architecture oran-architecture-v15 with a split distributed design in the 1800 MHz band (band 3), where four gNBs (5G base stations) provide radio access to user equipment (UE). Each gNB includes a radio unit (RU) that is connected to a distributed unit (DU) via a 100 Mbit/s fronthaul (the Uu interface). DUs are aggregated into two virtual baseband units (vBBUs), each comprising one centralized unit (CU) and two DUs. They are interconnected through a 1 Gbit/s midhaul (the F1 interface) and hosted in an edge cloud together with a near-real-time RAN intelligent controller (near-RT RIC) that interfaces with a non-RT RIC (the A1 interface), which is deployed in a service management and orchestration (SMO) platform. The vBBUs are interconnected through the Xn interface and communicate with the core network over a 1 Gbit/s backhaul (the NG interface). The core network follows a service-oriented architecture and provides (virtual) network functions, such as the access and mobility management function (AMF). It is connected to the data network (e.g., the Internet) via the N6 interface.
  • Figure 3: Causal (summary) graph that represents the causal structure of the cloud radio access network in Fig. \ref{['fig:target_system']}. For brevity, we use plate notation to represent the structure of sets of variables plate_notation, i.e., the causal structure in the left box applies to the four distributed units (DUs) and the causal structure in the right box applies to the two central units (CUs). In total, the graph has $220$ nodes (system variables) and $367$ directed edges (causal dependencies). UL is an acronym for uplink and DL is an acronym for downlink. The SNR refers to the signal-to-noise ratio measured on the uplink of a DU. The MCS represents the average modulation and coding scheme across the uplink and downlink of a DU. Similarly, the BER represents the block error rate across the uplink and downlink of a DU. TX is a shorthand for transmitter and RX is a shorthand for receiver. Jitter is the variation in the delay (latency) of data packets.
  • Figure 4: Schematic of our method for identifying the safe region $\mathcal{S}_{\mathbf{U}'_{t_k}}$ of control inputs: $\mathbf{U}'_{t_k} = \{U_{t_k},V_{t_k}\}$. We first observe the RAN for $t_0$ time steps to infer an initial safe region $\hat{\mathcal{S}}_{\mathbf{U}'_{t_0}}$ via causal inference, after which we gradually expand the estimated region through a sequence of interventions.
  • Figure 5: Example networked system [cf. a)] and its causal structure [cf. b)]. The system consists of edge servers that handle service requests from UEs.
  • ...and 9 more figures

Theorems & Definitions (6)

  • Remark 1
  • Definition 1: Specification
  • Definition 2: Safe region
  • Definition 3: Causal effect identifiability
  • Proposition 1
  • Proposition 2