PaperlayerView exact revision on arXiv

Source-linked AI summary

Caching in the Sky: Proactive Deployment of Cache-Enabled Unmanned Aerial Vehicles for Optimized Quality-of-Experience

Mingzhe Chen, Mohammad Mozaffari, Walid Saad, Changchuan Yin, Mérouane Debbah, Choong-Seon Hong

arXiv:1610.01585v1cs.IT

TL;DR

The paper studies proactive cache-enabled UAV deployment for serving mobile users while meeting QoE requirements with minimum transmit power. It leverages user-centric information and echo state networks to support deployment decisions, reporting a 40% gain in average transmit power.

  • Problem

    The paper addresses how to serve mobile users while guaranteeing each user's QoE requirement using minimum UAV transmit power.

  • Method

    The proposed framework uses user-centric content-request and mobility information with an echo state network algorithm to deploy flying UAV services.

  • Results

    40% gain in average transmit power is reported for the proposed algorithm.

  • Takeaways & Limitations

    The framework combines flying UAV service with predictive, user-centric deployment to improve caching efficiency and transmit-power performance.

Abstract

from arXiv · show

In this paper, the problem of proactive deployment of cache-enabled unmanned aerial vehicles (UAVs) for optimizing the quality-of-experience (QoE) of wireless devices in a cloud radio access network (CRAN) is studied. In the considered model, the network can leverage human-centric information such as users' visited locations, requested contents, gender, job, and device type to predict the content request distribution and mobility pattern of each user. Then, given these behavior predictions, the proposed approach seeks to find the user-UAV associations, the optimal UAVs' locations, and the contents to cache at UAVs. This problem is formulated as an optimization problem whose goal is to maximize the users' QoE while minimizing the transmit power used by the UAVs. To solve this problem, a novel algorithm based on the machine learning framework of conceptor-based echo state networks (ESNs) is proposed. Using ESNs, the network can effectively predict each user's content request distribution and its mobility pattern when limited information on the states of users and the network is available. Based on the predictions of the users' content request distribution and their mobility patterns, we derive the optimal user-UAV association, optimal locations of the UAVs as well as the content to cache at UAVs. Simulation results using real pedestrian mobility patterns from BUPT and actual content transmission data from Youku show that the proposed algorithm can yield 40% and 61% gains, respectively, in terms of the average transmit power and the percentage of the users with satisfied QoE compared to a benchmark algorithm without caching and a benchmark solution without UAVs.

I. INTRODUCTION

The paper addresses mobile-user caching in CRANs by combining human-centric behavior prediction with cache-enabled UAV deployment. It jointly considers QoE, transmit power, associations, UAV locations, and cached contents, reporting gains over benchmarks.

  • Research gap: Prior mobility-prediction studies focused on prediction rather than using mobility patterns for caching and wireless resource optimization.Existing caching works were also typically restricted to static users and terrestrial base stations.
  • Research gap: Static terrestrial caches may fail to serve mobile users outside a base station’s coverage or when requested content is unavailable in a new cell.Replicating content across multiple base stations can serve such users but is described as inefficient.
  • Proposed direction: Cache-enabled UAVs can dynamically cache popular content, track corresponding users’ mobility, and directly transmit stored content, reducing fronthaul traffic.Caching can proactively download content during off-peak hours or while UAVs are at docking stations.
  • Framework: The framework uses user-centric content-request and mobility information to deploy cache-enabled UAVs while maximizing QoE with minimum total UAV transmit power.Its QoE metric incorporates transmission delay and device-dependent perceptions of rate requirements.
  • Framework: A conceptor-based ESN separates users’ behaviors into patterns learned independently, and the framework derives user-UAV associations, UAV locations, and cached content for mobile users in CRANs.The paper positions this as extending prior UAV deployment studies that assumed static users and did not study UAV caching.
  • Evaluation: 40% gain in average transmit power is reported against a baseline without cache, while 61% gain in users with satisfied QoE is reported against a benchmark without UAVs.The evaluation uses Youku content-request data and measured BUPT mobility data.

II. SYSTEM MODEL AND PROBLEM FORMULATION

The CRAN deploys cache-enabled UAVs alongside terrestrial RRHs to serve mobile users, using predicted mobility and requests to guide associations, placement, and caching. The model specifies communication, mobility, caching-refresh, and transmission-time assumptions.

  • II. SYSTEM MODEL AND PROBLEM FORMULATION: The CRAN serves mobile users through terrestrial RRHs and K cache-enabled UAVs acting as flying cache-enabled RRHs.RRHs are grouped into BBU clusters, while UAVs provide additional service alongside the terrestrial infrastructure.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: UAV-user links use mmWave, while terrestrial RRHs use the cellular band and connect to the cloud’s BBU pool.UAV altitude reduces obstacle blocking, whereas the terrestrial infrastructure provides the conventional access path.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: UAV caches store popular predicted contents so UAVs can transmit directly to users, reducing content-server-to-UAV delay and helping when RRHs cannot satisfy QoE.Unlike RRH caching, UAV caching supports users whose associations and locations change with mobility.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: Each user requests at most one equal-sized content per time slot, and UAV caches refresh every T slots during off-peak docking periods.The transmission duration Δτ is algorithmically determined and bounds each content transmission.
  • A. Mobility Model: Users follow periodic pedestrian mobility patterns inferred from locations collected every H time slots, with constant-speed movement between observed locations.Examples include repeatedly visiting a workplace at similar times on weekdays.
  • A. Mobility Model: UAV associations may change with QoE, UAVs remain static during each transmission, and their locations update after transmission according to user mobility.The predicted mobility pattern determines cached contents and UAV locations, affecting user QoE.

B. Transmission Model

The transmission model represents UAV, fronthaul, and RRH-user links with time-discretized channel, interference, bandwidth, and capacity models. It accounts for probabilistic LoS/NLoS propagation and ZFBF-based RRH transmission.

  • B. Transmission Model: Each slot is divided into F small intervals so user locations remain approximately constant during an interval.The relation is Δτ = Ft.
  • 1) UAVs-Users Links: UAV-user propagation models LoS and NLoS path loss, shadowing, elevation angle, environment-dependent LoS probability, and average SNR.The model uses mmWave channel parameters and log-normal shadowing.
  • 1) UAVs-Users Links: The UAV-user channel capacity is determined from the average SNR while each UAV’s bandwidth BV is equally divided among its associated users.The resulting per-user capacity is denoted CV.
  • 2) BBUs-UAVs Ground-to-Air Links: BBU-UAV links use probabilistic LoS/NLoS propagation over the licensed cellular band, which is modeled as more reliable than mmWave for longer fronthaul distances.NLoS links experience higher attenuation than LoS links.
  • 3) RRHs-Users Links: RRHs are grouped into clusters and use zero-forcing beamforming to serve associated users, with received signals including inter-cluster and wireless-fronthaul interference.The RRH model specifies channel gains, beamforming vectors, transmit power, bandwidth, and noise.
  • 3) RRHs-Users Links: RRH-user capacity is derived from the received SINR using cluster channel matrices, beamforming matrices, transmitted content, and noise power.The model assumes equal RRH transmit power and assigns each associated user bandwidth B.

C. Quality-of-Experience Model

The QoE model combines data rate, delay, and device type, while modeling delivery over RRH, UAV fronthaul, and UAV-cache paths. Delay sensitivity is mapped to MOS categories, and UAV power can be adjusted to meet QoE requirements.

  • C. Quality-of-Experience Model: QoE is a human-in-the-loop metric that captures each user’s data rate, delay, and device type.The model therefore evaluates service quality using more than throughput alone.
  • 1) Delay: Content reaches users through content server-BBUs-RRHs-user, content server-BBUs-UAV-user, or UAV cache-user links.The backhaul from the cloud to the core network is assumed to use fiber, so its delay is neglected.
  • 1) Delay: RRH-served users share the wired fronthaul rate, giving each user vF/NFR when NFR users receive contents from RRHs.This per-user rate contributes to the delay calculation.
  • 1) Delay: The delay model combines link delays for each content transmission and uses the relevant transmission rates for RRH, UAV fronthaul, and UAV-user paths.The UAV fronthaul rate is calculated analogously to the UAV-user rate model.
  • 1) Delay: Proposition 1 gives a lower delay bound based on maximum UAV transmit power, minimum UAV altitude, and the rates of the transmission links.The proof compares the three link types and identifies the minimum delay of link (a) as L/vF.
  • 1) Delay: Increasing user load lowers fronthaul rates, so UAV transmit-power adjustment can be used to satisfy QoE requirements within the system’s delay bound.Delay sensitivity is categorized into five groups using the MOS model.

2) Device Type:

The device type, represented through screen size, affects users’ QoE by shaping required data rates and device-related MOS. The formulation uses these requirements alongside delay, associations, caching, UAV locations, and transmit power to optimize service.

  • Device-size modeling: Larger-screen devices are modeled as more QoE-sensitive because they typically display higher-resolution content and require higher data rates.The device diameter is represented by S_i, which scales the content-specific rate requirement.
  • Device-size modeling: δS_i,n = S_i ˆC_n defines the rate requirement for user i with device size S_i receiving content n.ˆC_n is the rate requirement associated with content n.
  • QoE mapping: The device type score can be 1 or 0, corresponding to MOS values of “Excellent” or “Poor”.The device-type component is mapped into the MOS-based QoE model.
  • QoE mapping: For 0.8 ≤ D̄τ,i,n ≤ 1, the delay MOS is “Excellent”, with D̄min = 0.8 minimizing delay.The delay rate requirement is defined as the rate that achieves the optimal delay.
  • Optimization formulation: The deployment problem jointly predicts user behavior, selects cached contents and user-UAV associations, places UAVs, and minimizes transmit power while enhancing QoE.Predicted mobility informs UAV locations, while predicted content-request distributions inform cached content selection.

III. CONCEPTOR ECHO STATE NETWORKS FOR CONTENT AND MOBILITY PREDICTIONS

The paper proposes conceptor-based ESNs to predict user-specific content-request distributions and mobility patterns from contextual information. Conceptors let the ESNs represent multiple nonlinear patterns and add new patterns without interfering with previously learned ones.

  • Overview: The proposed prediction algorithm uses ESNs with conceptors to predict users’ content-request distributions and mobility patterns.The approach is designed for behavior prediction using limited state information.
  • Inputs and outputs: Each user’s context includes time-related and demographic or device attributes such as gender, occupation, age, and device type.The context is used to predict corresponding content-request and mobility behavior.
  • Motivation: ESNs can learn mobility patterns and content-request distributions quickly without significant training data because of the echo state property.This is contrasted with traditional neural-network and deep-learning approaches.
  • Conceptor role: Conceptors enable multiple mobility and content-request patterns to be learned and allow new patterns to be added without interfering with earlier ones.They characterize the ESN reservoir and support different nonlinear systems.
  • Inputs and outputs: For content prediction, the ESN output is a probability vector representing the likelihood that a user requests each content.The ESN models the relationship between user context and content-request distribution.

2) Mobility pattern prediction:

The mobility-prediction component uses conceptor ESNs to map each user’s current location and context to predicted future locations. Daily mobility during a week is treated as a distinct prediction pattern.

  • Model structure: The mobility-prediction ESN model uses an output weight matrix and a recurrent matrix.BBUs implement user-specific conceptor ESN algorithms for mobility prediction.
  • Mobility inputs and outputs: The mobility ESN takes a user’s current location and context as input and predicts the user’s future locations.The output contains predicted locations over the next Ns time duration H.
  • Prediction patterns: Each user’s mobility during each day of one week is treated as one prediction pattern.The conceptor expression is the same as in the content-request prediction case.
  • Training: The algorithm trains conceptors and output weights using users’ contexts, corresponding content requests, and reservoir states.The training stage combines updated reservoir states from prediction patterns and uses ridge regression for the output matrix.

2) Prediction Stage:

In the prediction stage, conceptors control reservoir updates so one ESN architecture can generate different learned behavior patterns. The resulting predictions are then used to support cache-enabled UAV deployment decisions.

  • Prediction stage: Changing the conceptor controls reservoir-state updates and allows different patterns to be predicted within one ESN architecture.The content-request distribution prediction is obtained from the controlled reservoir state and output model.
  • Prediction stage: The input simulation matrix controls the ESN reservoir’s memory during prediction.The algorithm uses this mechanism to manage learned pattern information.
  • Prediction stage: The proposed conceptor ESN learns each prediction pattern through a unique nonlinear system.This enables behavior prediction with different nonlinear systems during different time periods.
  • Deployment decisions: The predicted content distributions and mobility patterns are used to determine UAV associations, cached contents, and UAV locations.Users are clustered for UAV service, and each UAV serves one cluster in the described model.

A. Users-RRH Association

The framework determines user-RRH associations from predicted locations and rate requirements, then uses associations and content-request distributions to select cached contents.

  • User-RRH association depends on each user's fronthaul rate, delay-rate requirement, and device-rate requirement.
  • As more users join an RRH cluster, each user's fronthaul rate decreases, improving the delay interval.
  • Users are clustered by proximity so each UAV services one cluster, determining the user-UAV association.
  • Caching reduces transmission delay and the UAV transmit power needed to satisfy users' QoE requirements.
  • The optimal cached contents depend on prior user associations and each user's predicted content-request distribution.

C. Optimal Locations of UAVs

The paper derives UAV locations for two special altitude cases and uses learning to obtain power-efficient sub-optimal locations in more general cases.

  • The location optimization assigns each UAV to serve its associated users while meeting their QoE-related rate requirements.
  • Theorem 3 gives optimal UAV locations for two special cases to minimize each UAV's transmit power.
  • Closed-form location derivations become highly challenging in generic cases because altitude depends on the UAV's x and y coordinates.
  • A learning algorithm adapts UAV locations to network states and produces sub-optimal, power-efficient service locations.
  • The centralized implementation lets BBUs determine UAV locations without information exchange and directly adjust locations or associations when predictions are inaccurate.

V. SIMULATION RESULTS

Simulations evaluate prediction behavior, transmit power, QoE, and UAV placement using pedestrian mobility and content-request data. The proposed approach improves power and satisfied-QoE performance across comparisons.

  • Conceptor ESN memory increases with learned mobility patterns and uses less memory for pattern 2 because it resembles pattern 1.
  • 40% and 25% gains in average transmit-power reduction occur versus the proposed algorithm without cache and without UAV-location optimization, respectively.
  • Caching can reduce the rate required to reach users' QoE thresholds when wireless fronthaul rates are low.
  • 61% gain in the percentage of users with satisfied QoE occurs versus the proposed algorithm without UAVs.
  • 85% lower average UAV transmit power results when the number of UAVs increases from 3 to 7.
  • Theorem 3's optimal-location approximation has only 5.8% deviation from exhaustive search, while its solution approaches optimality as UAV count increases.

VI. CONCLUSIONS

The paper formulates QoE-constrained UAV deployment as a minimum-transmit-power problem and solves it with conceptor-based echo state networks that predict user behavior from limited information.

  • The framework uses flying UAVs to serve mobile users in a CRAN system.
  • The optimization guarantees each user's QoE requirement while minimizing UAV transmit power.
  • The proposed algorithm uses echo state networks and conceptors to predict content-request distributions with limited network-state and user-context information.
  • The ESNs separate user behavior into several patterns and learn those patterns with nonlinear systems.
  • Simulation results show significant gains in minimum transmit power compared with conventional ESN approaches.
Loading 1610.01585v1…