PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Cell LTE-U/Wi-Fi Coexistence Evaluation Using a Reinforcement Learning Framework

José M. de C. Neto, Sildolfo F. G. Neto, Pedro M. de Santana, Vicente A. de Sousa Jr

arXiv:2610.20230v1cs.NI

TL;DR

Growing broadband IoT demand strains licensed spectrum, while Wi-Fi's established use of the unlicensed band makes coexistence important. The paper evaluates multi-cell LTE-U/Wi-Fi coexistence and proposes centralized, coordinated Q-Learning to adjust LTE-U duty cycle. The framework improves aggregate and user data rates in the evaluated setting, including gains over fixed duty-cycle choices.

  • Problem

    Growing broadband IoT data demand increases pressure on scarce licensed spectrum, while LTE-U's use of unlicensed spectrum must coexist with established Wi-Fi.

    Source: §1 · pp. 1–2

  • Method

    A centralized, coordinated Q-Learning framework adapts LTE-U duty cycle using transmitted data rates in a multi-cell coexistence scenario.

    Source: §1 · p. 2, §5.2 · p. 13

  • Results

    Around 3 Mbps more maximum aggregated transmitted data rate than the best static DC values of 0.6 and 0.7, a 5% increase; gains extend to system and user transmitted data rates.

    Source: §6 · p. 15, §7 · p. 18

  • Takeaways & Limitations

    In the evaluated setting, centralized coordination places the Q-Learning energy cost at the central node and improves data rate without added energy consumption for each terminal.

    Source: §7 · p. 18

    Same-access-technology interference is not explicitly modeled, although its packet losses are reflected in collected data rates.

    Source: §7 · p. 18

Abstract

from arXiv · show

Cellular broadband Internet of Things (IoT) applications are expected to keep growing year-by-year, generating demands from high throughput services. Since some of these applications are deployed over licensed mobile networks, as long term evolution (LTE), one already common problem is faced: the scarcity of licensed spectrum to cope with the increasing demand for data rate. The LTE-Unlicensed (LTE-U) forum, aiming to tackle this problem, proposed LTE-U to operate in the 5 GHz unlicensed spectrum. However, Wi-Fi is already the consolidated technology operating in this portion of the spectrum, besides the fact that new technologies for unlicensed band need mechanisms to promote fair coexistence with the legacy ones. In this work, we extend the literature by analyzing a multi-cell LTE-U/Wi-Fi coexistence scenario, with a high interference profile and data rates targeting a cellular broadband IoT deployment. Then, we propose a centralized, coordinated reinforcement learning framework to improve LTE-U/Wi-Fi aggregate data rates. The added value of the proposed solution is assessed by a ns-3 simulator, showing improvements not only in the overall system data rate but also in average user data rate, even with the high interference of a multi-cell environment.

1. Introduction

Growing broadband IoT demand strains licensed spectrum, while LTE-U seeks additional unlicensed capacity in a band already used by Wi-Fi. This paper extends its coexistence framework to multiple cells and uses coordinated reinforcement learning to improve aggregate data rates.

1. Introduction

  • Licensed spectrum scarcity and growing mobile-data demand motivate using unlicensed spectrum for broadband IoT, but Wi-Fi already operates there and coexistence must be considered.The paper cites multimedia, augmented-reality, and virtual-reality services as examples of data-intensive IoT applications.Source: §1 · pp. 1–2
  • LTE-U targets the 5 GHz unlicensed band in regions without mandatory LBT and uses carrier aggregation and a TDM-like duty cycle for coexistence.The duty cycle creates ON/OFF periods and may reduce energy consumption.Source: §1 · p. 2
  • LAA uses a licensed anchor and LBT for regions where it is mandatory, while MuLTEfire is designed for stand-alone unlicensed operation.These are among the three proposed approaches to unlicensed LTE operation described in the introduction.Source: §1 · p. 2
  • Wi-Fi and LTE-U/LAA share 5 GHz and use OFDM, but their different MAC layers motivate coexistence investigation for broadband applications.Wi-Fi is based on IEEE 802.11, while the paper identifies MAC differences as a reason to study coexistence.Source: §1 · p. 2
  • The paper extends prior Q-Learning duty-cycle adaptation to co-channel multi-cell LTE-U/Wi-Fi coexistence, using transmitted data rate to assess interference and improve aggregate rate.The proposed framework is centralized and coordinated.Source: §1 · p. 2

2. Related Works

Prior coexistence work spans access mechanisms, optimization, and machine learning, with reported gains across different metrics and settings. This paper extends its earlier Q-Learning approach to a coordinated, centralized multi-cell evaluation with multiple users and more challenging interference.

2. Related Works

  • Classical optimization and game-theoretic approaches address throughput, delay, fairness, and QoS, with a reported trade-off between QoS and solution complexity.The cited work also includes threshold adjustment, resource sharing, and coalition or bargaining games.Source: §2 · p. 3
  • Related work covers carrier sensing, link adaptation, power control, spectrum slicing, and channel selection as strategies for managing coexistence.The surveyed approaches target throughput, fairness, QoS, or interference through different radio-resource-management choices.Source: §2 · pp. 3–4
  • Cooperative coexistence approaches exchange messages or centrally manage resources, while machine learning is identified as suitable for this problem.Prior work includes negotiated operating modes and centralized control of distributed access points.Source: §2 · pp. 3–4
  • Earlier Q-Learning work dynamically adapted LTE-U duty cycle but evaluated only one device and one cell per operator, motivating a more realistic multi-cell study.The new evaluation adds devices and cells to make the interference profile more realistic and challenging to track.Source: §2 · p. 5
  • The paper contributes a centralized Q-Learning mechanism that adjusts duty cycle under co-channel interference and distinguishes user from operator data rates.Its evaluation covers multi-cell coexistence with changing offered data rates and co-channel and inter-RAT interference.Source: §2 · p. 5

3. Wi-Fi and LTE-U MAC Layer

The section reviews Wi-Fi's distributed contention-based access and LTE-U's layered coexistence mechanisms. Wi-Fi stations contend for channel access, while LTE-U combines carrier and channel choices with sensing-driven duty cycling.

3.1. The 802.11 MAC Layer

  • Wi-Fi's DCF is a distributed LBT contention mechanism in which a station senses the channel and defers if it becomes busy before transmission.Stations use DIFS and, when necessary, a random backoff before transmitting.Source: §3.1 · p. 6
  • Random backoff reduces collision probability when multiple Wi-Fi stations are waiting as the channel becomes idle.A station selects a random integer in the contention-window range after the channel is idle for DIFS.Source: §3.1 · p. 6
  • Wi-Fi checks medium occupancy through physical carrier sensing and virtual carrier sensing, with the latter using the MAC-header duration to set NAV.Physical sensing uses preamble detection and energy measurements.Source: §3.1 · p. 6

3.2. The LTE-U Coexistence Mechanisms

  • LTE-U's coexistence mechanisms are supplemental downlink, channel selection, and carrier-sense adaptive transmission (CSAT).The licensed PCell carries control and signaling, while an opportunistically enabled unlicensed SCell boosts downlink capacity.Source: §3.2 · pp. 6–7
  • LTE-U selects a cleaner channel using interference measurements; if none is clear, CSAT senses the medium and proportionally activates or deactivates the SCell by duty cycle.During deactivated periods, the channel is released for coexisting technologies; ABS can implement the selected blank-subframe proportion.Source: §3.2 · pp. 7–8

4. System model, Evaluation Scenario and Preliminary Results

The section defines a challenging 3GPP indoor multi-cell evaluation and shows that the duty cycle maximizing aggregate data rate varies with offered load, motivating automatic selection.

4.1. System Model and Evaluation Scenario

  • The scenario is a 3GPP indoor hotspot with two operators, four small cells each, and 40 stationary stations in a 50 × 120 m room.The room has no walls, and stations are randomly distributed.Source: §4.1 · p. 8, Figure 3 · p. 9
  • The evaluation uses ns-3 with LTE-U duty-cycle transmission based on a 40 ms ABS pattern, controlled by a 40-bit subframe mask.Each LTE subframe lasts 1 ms.Source: §4.1 · p. 9
  • Duty cycles from 0.2 to 0.9 are tested under full-buffer constant-bit-rate UDP traffic, with simulations lasting 20 seconds.The offered data rate represents available capacity per station; the setup aims to sustain contention for medium access.Source: §4.1 · p. 9

4.2. Preliminaries Results

  • At 2 Mbps offered per station, LTE-U plus Wi-Fi aggregate transmitted rate peaks at DC 0.6; each operator nearly reaches its 40 Mbps interference-free expectation at some duty cycles, but not simultaneously.Source: Figure 4 · p. 10, §4.2 · pp. 10–11
  • At 4 Mbps per station, aggregate transmitted rate peaks at DC 0.4, with Wi-Fi having more channel access time than LTE-U.The operator data rates also vary more with duty cycle than at 2 Mbps per station.Source: Figure 5 · p. 11, §4.2 · p. 11
  • Because the maximizing duty cycle depends on offered data rate, dynamic-load operation requires automatic on-the-fly duty-cycle selection.The section frames this as a sequential decision problem addressed in the following reinforcement-learning section.Source: §4.2 · p. 11

5. Proposed Reinforcement Learning Framework

The paper frames duty-cycle selection as a sequential decision problem and applies Q-Learning to coordinate LTE-U cells toward maximizing aggregate transmitted data rate. The framework uses observed rates as states and rewards, with a central node selecting and distributing duty-cycle actions.

5. Proposed Reinforcement Learning Framework

  • Q-Learning is used because changing offered data rates make duty-cycle selection a sequential decision problem in which each choice leads to a new system-rate state.The paper seeks a duty cycle that maintains aggregate data rate as offered rates change.Source: §5 · p. 11

5.1. Q-Learning

  • Q-Learning estimates action values from rewards without requiring the transition probabilities needed to solve for the optimal value directly.The paper states convergence with probability 1 if every state–action pair continues to be updated.Source: §5.1 · pp. 12–13

5.2. Proposed Framework

  • The framework centrally selects a duty cycle for all LTE-U cells to maximize system aggregate transmitted data rate.The selected value is sent to the cells at each step.Source: §5.2 · pp. 13–14
  • The Q-Learning algorithm can choose duty-cycle actions A = {0.2, 0.4, 0.6, 0.8}; its four states partition aggregated Wi-Fi and LTE-U transmitted data rates into intervals from 0 to M.These actions represent four airtime situations, from Wi-Fi having more airtime at DC = 0.2 to LTE-U having more at DC = 0.8.Source: §5.2 · p. 13
  • The reward is aggregate transmitted data rate, and the algorithm updates the shared duty cycle through repeated cell reporting and central Q-Learning cycles.The target rate Ttarget is a desired parameter, distinct from M, the maximum achievable rate under the system parameters.Source: §5.2 · pp. 13–14

1 Initialize

The algorithm initializes its Q-table with random Q-values.

1 Initialize

  • The algorithm initializes the Q-table with random Q-values Q(s, a).Source: §1 · p. 14

4 end

The algorithm estimates initial states before learning.

4 end

  • Before learning, the algorithm estimates the initial states s.Source: §4 · p. 14

13 end

The algorithm takes an immediate reward and updates its Q-table.

13 end

  • The algorithm takes immediate reward r as Ttx wi f i + Ttx wi f i, then updates the Q-table according to the stated expression.Source: §13 · p. 14

6. Proposed Solution Evaluation

In 250-second simulations with changing offered loads and varied station positions, Q-Learning selected duty cycles every 40 ms. It improved maximum aggregate throughput over the best static settings, while user-level outcomes depended on the static duty cycle used for comparison.

6. Proposed Solution Evaluation

  • The 250-second simulations used 100 station-placement snapshots, offered rates of 500 kbps to 4 Mbps, and Q-Learning duty-cycle updates every 40 ms.Offered data rates changed every 10–15 seconds; the Q-Learning parameters were α = 0.3 and γ = 0.5.Source: §6 · p. 15
  • Around 3 Mbps (5%) higher maximum aggregated transmitted data rate than the best static duty cycles, DC = 0.6 and DC = 0.7, was achieved with Q-Learning.The system-level objective was to maximize the combined LTE-U and Wi-Fi transmitted data rate, not either operator’s individual maximum.Source: §6 · p. 15, Figure 8 · p. 16
  • Against static DC = 0.6, Q-Learning improved transmitted data rates for practically all Wi-Fi users while LTE-U users maintained their performance.The approximately 3 Mbps aggregate gain exceeded the 2.7 Mbps maximum user rate and was distributed among users with lower data rates.Source: §6 · p. 16, Figure 9 · p. 16
  • Against static DC = 0.7, Q-Learning increased Wi-Fi rates but worsened LTE-U performance, especially for LTE-U stations with fewer data rates.The reported aggregate gain of 3 Mbps was distributed across practically all Wi-Fi stations.Source: Figure 10 · p. 17, §6 · p. 17
  • At the 10th percentile, Wi-Fi Q-Learning rates exceeded static duty cycles above 0.5 while LTE-U rates exceeded those below 0.5; at the 90th percentile, the corresponding thresholds were above 0.4 for Wi-Fi and below 0.5 for LTE-U.At the 90th percentile, LTE-U rates varied little across high duty-cycle values.Source: Figure 11 · p. 17, Figure 12 · p. 18, §6 · p. 18

7. Conclusions and Future Works

The paper evaluates centralized, coordinated Q-Learning for multi-cell LTE-U/Wi-Fi coexistence with same- and different-technology interference. The 40 ms duty-cycle selection improves aggregate and user transmitted data rates; the model does not explicitly account for same-technology interference.

7. Conclusions and Future Works

  • The scenario includes interference between the same and different access technologies, and the maximum data rate varies with duty-cycle value and offered user data rate.The authors describe this multi-cell setting as more challenging than previous work.Source: §7 · p. 18
  • Duty-cycle selection every 40 ms improved aggregate transmitted data rate, with gains distributed across Wi-Fi users and improvements in system and user rates.The framework uses centralized, coordinated Q-Learning in a multi-cell LTE-U/Wi-Fi coexistence scenario.Source: §7 · p. 18
  • For multimedia IoT deployments, centralizing Q-Learning confines its additional energy consumption to the central node, avoiding added energy cost for each terminal while improving data rate.The authors frame this as a trade-off between energy consumption and increasing data rate.Source: §7 · p. 18
  • Same-technology interference is not explicitly modeled, though its effects are included in collected throughput values through lost packets.Future work may explicitly measure interference and explore QoS modifications or a decentralized approach.Source: §7 · pp. 18–19
Loading 2610.20230v1…