PaperlayerView exact revision on arXiv

Source-linked AI summary

Liquid State Machine Learning for Resource and Cache Management in LTE-U Unmanned Aerial Vehicle (UAV) Networks

Mingzhe Chen, Walid Saad, Changchuan Yin

arXiv:1801.09339v1cs.IT

TL;DR

The paper studies joint caching and resource allocation for cache-enabled LTE-U UAV networks serving ground users under limited cloud–UAV fronthaul information. It uses a distributed liquid state machine to predict content requests and select resource strategies, achieving up to 33.3% and 50.3% gains in stable-queue users over Q-learning baselines.

  • Problem

    The central problem is jointly optimizing user association, licensed/unlicensed spectrum allocation, and content caching for LTE-U UAV networks serving ground users.

  • Method

    A distributed liquid state machine predicts users’ content-request distributions and enables UAVs to autonomously choose resource allocation strategies from limited network-state information.

  • Results

    33.3% and 50.3% gains in stable-queue users are reported over Q-learning with cache and Q-learning without cache, respectively.

  • Takeaways & Limitations

    The approach derives UAV cache contents and resource allocations from user associations and predicted content-request distributions.

Abstract

from arXiv · show

In this paper, the problem of joint caching and resource allocation is investigated for a network of cache-enabled unmanned aerial vehicles (UAVs) that service wireless ground users over the LTE licensed and unlicensed (LTE-U) bands. The considered model focuses on users that can access both licensed and unlicensed bands while receiving contents from either the cache units at the UAVs directly or via content server-UAV-user links. This problem is formulated as an optimization problem which jointly incorporates user association, spectrum allocation, and content caching. To solve this problem, a distributed algorithm based on the machine learning framework of liquid state machine (LSM) is proposed. Using the proposed LSM algorithm, the cloud can predict the users' content request distribution while having only limited information on the network's and users' states. The proposed algorithm also enables the UAVs to autonomously choose the optimal resource allocation strategies that maximize the number of users with stable queues depending on the network states. Based on the users' association and content request distributions, the optimal contents that need to be cached at UAVs as well as the optimal resource allocation are derived. Simulation results using real datasets show that the proposed approach yields up to 33.3% and 50.3% gains, respectively, in terms of the number of users that have stable queues compared to two baseline algorithms: Q-learning with cache and Q-learning without cache. The results also show that LSM significantly improves the convergence time of up to 33.3% compared to conventional learning algorithms such as Q-learning.

I. INTRODUCTION

The paper addresses joint caching and resource allocation for cache-enabled LTE-U UAV networks under cloud–UAV fronthaul constraints. It proposes a liquid state machine framework that jointly supports caching, user association, and licensed/unlicensed spectrum allocation.

  • Motivation: Caching popular contents at UAVs can offload fronthaul traffic by allowing direct transmission to ground users.Direct delivery avoids wireless fronthaul transmissions whenever possible.
  • Motivation: LTE-U capabilities allow UAVs to use both licensed and unlicensed spectrum to serve ground users.The paper combines this spectrum flexibility with UAV-side caching.
  • Contributions: The proposed framework jointly analyzes licensed and unlicensed resource allocation, user association, and cache-content replacement to maximize stable-queue users.The optimization objective is the number of users with stable queues.
  • Contributions: The LSM approach jointly performs caching and resource allocation while accounting for limited-capacity UAV–cloud links.The paper positions this combination as addressing a gap in prior UAV and LTE-U studies.
  • Contributions: The cloud predicts users’ content-request distributions, while UAVs autonomously adjust spectrum allocation using limited network-state information.The predicted distributions support content-cache decisions at individual UAVs.
  • Simulation results: 33.3% and 50.3% gains in stable-queue users are achieved over Q-learning with cache and without cache, respectively.The comparison concerns the number of users having stable queues.

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system models cache-enabled UAV base stations serving LTE-U users through licensed and unlicensed bands, with WiFi coexistence and cloud-based content delivery. Its formulation combines association, caching, and resource allocation under explicit spectrum-sharing and storage assumptions.

  • II. SYSTEM MODEL AND PROBLEM FORMULATION: UAVs act as flying cache-enabled LTE-U base stations controlled by a cloud server and serving ground users over licensed and unlicensed bands.Each UAV can allocate at most one resource type to each user, while cloud–UAV traffic uses licensed wireless fronthaul.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: Licensed LTE uses FDD downlink operation, while LTE-U uses TDD duty cycling to divide unlicensed slots between LTE-U and WiFi users.LTE-U transmits for fraction ϑ and is muted for fraction 1 −ϑ allocated to WiFi.
  • A. WiFi data rate analysis: WiFi access points use CSMA/CA with distributed coordination and RTS/CTS mechanisms, sharing the unlicensed band with LTE-U.The model assumes slotted exponential backoff and one LTE slot containing TW WiFi slots.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: The optimization jointly considers licensed and unlicensed resource allocation, user association, and cache-content replacement to maximize stable-queue users.The contents are stored at a cloud-based server and requested by users as equal-sized objects.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: User content requests follow probability vectors pi, and the cloud obtains these distributions through the learning algorithm described later.Each UAV’s storage holds C popular contents from the requestable content set.
  • II. SYSTEM MODEL AND PROBLEM FORMULATION: Each UAV stores a set of popular contents, and suitable caching can offload fronthaul traffic through direct UAV–user transmission.Cached contents are periodically refreshed during off-peak hours when UAVs return to docking stations.
  • A. WiFi data rate analysis: The LTE-U time fraction is constrained by WiFi demand as ϑ ≤1 −Nwγ/R(Nw), where γ is each WiFi user’s rate requirement.This constraint ensures the WiFi rate requirement is satisfied.

B. UAV data rate analysis

The paper derives UAV-user and cloud–UAV rates across licensed and unlicensed links using path-loss, bandwidth, transmission-power, and allocation models. These rates distinguish cached direct delivery from cloud-served traffic over capacity-shared fronthaul.

  • B. UAV data rate analysis: UAV-user transmission combines wireless fronthaul and air-to-ground links, with probabilistic LoS/NLoS propagation over the licensed band.NLoS links experience higher attenuation because of shadowing and diffraction loss.
  • 1) UAVs-users links over the licensed band: LoS probability depends on environment, building density and height, user/UAV locations, and elevation angle.The model uses environment-dependent constants in the LoS probability expression.
  • 1) UAVs-users links over the licensed band: The licensed-band downlink rate depends on bandwidth, UAV and cloud transmit powers, channel gain, noise, path loss, and allocation fraction uki(t).The licensed allocation fractions satisfy a sum-to-one constraint across users associated with a UAV.
  • 2) UAVs-users links over the unlicensed band: The unlicensed-band UAV-user rate incorporates the UAV time fraction ϑ, unlicensed bandwidth, and average unlicensed-band path loss.Unlicensed allocation indicators also satisfy a sum-to-one constraint.
  • 3) Cloud-UAVs ground-to-air links: Cloud–UAV fronthaul rates use the average cloud–UAV path loss and divide total fronthaul bandwidth equally among users receiving cloud contents.The number of such users is represented by UC(t).
  • 3) Cloud-UAVs ground-to-air links: Caching avoids cloud retrieval for stored contents, whereas cloud-requested contents traverse the cloud–UAV fronthaul before reaching users.The fronthaul rate is therefore defined for users receiving contents from the cloud.

C. Queueing model

The queueing model tracks each user's queue evolution under four possible cloud/UAV transmission links and uses queue stability to characterize service feasibility.

  • Transmission links: The model includes UAV-user transmission over licensed or unlicensed bands and cloud-UAV-user transmission over either band.These four links determine the applicable user rate expressions.
  • Queue stability: A queue is rate stable when its service rate is at least its content arrival rate, Rki (t) ≥ Vi (t).Queue stability is used to measure users' content transmission delay.

D. Problem formulation

The paper formulates joint user association, licensed and unlicensed resource allocation, and UAV caching as an optimization problem that maximizes users with stable queues.

  • System goal: The formulation allocates cache-enabled UAV resources across licensed and unlicensed bands to satisfy every user's queue-stability requirement.The model explicitly targets appropriate bandwidth and time-slot allocation for the two bands.
  • Optimization objective: The objective is to maximize the number of users whose queues remain stable by jointly selecting association, spectrum resources, and cached contents.The decision variables include each UAV's associated users, licensed-band bandwidth indicators, unlicensed-band time-slot indicators, and cache contents.
  • Resource constraints: LTE-U users may occupy the unlicensed band only when the average data rate of each WiFi user exceeds its desired threshold.Additional constraints limit licensed bandwidth and unlicensed time slots to each UAV's available resources.
  • Algorithm context: The proposed LSM-based algorithm is illustrated as a component of the formulated resource-allocation framework.The supplied figure identifies the proposed LSM-based algorithm but does not provide further optimization results.

III. LIQUID STATE NETWORKS FOR CONTENT PREDICTION AND SELF-ORGANIZING RESOURCE ALLOCATION

The joint caching and allocation problem is combinatorial, non-convex, and difficult because UAVs may lack users' content-request information. The paper addresses these challenges with liquid state machine learning for request prediction and resource allocation.

  • Problem challenges: The problem is difficult because user association, spectrum allocation, and content caching are interdependent, while UAVs may not know which contents users will request.The formulation is described as combinatorial and non-convex, limiting conventional optimization approaches.
  • LSM approach: The proposed liquid state machine approach predicts users' content-request distributions and performs wireless resource allocation.The approach is introduced as a response to the optimization and incomplete-request-information challenges.
  • Adaptive operation: LSM stores users' behavioral information so the cloud can predict request distributions and adapt spectrum allocation to changing network states.The method is presented as using limited network-state information while tracking behavior over time.
  • Algorithm workflow: The algorithm is designed to predict content requests and solve the resource-allocation problem through a sequence of LSM components and processing steps.The paper introduces the components first and then explains the full prediction and optimization process.

A. LSM Components

The proposed LSM-based algorithm separates content-request prediction and resource allocation into function-specific designs built from five common components.

  • Common components: Each LSM-based algorithm consists of agents, input, output, a liquid model, and an output function.These five components define the common structure shown for the proposed algorithm.
  • Function-specific design: The paper designs separate function-specific components for predicting content-request distributions and allocating resources.The separation reflects the different functions required by the two tasks.

1) Content request distribution prediction:

The LSM predicts each user’s content-request distribution from contextual and dynamic user information. Its liquid model stores temporal features, while an offline-trained output function produces content-request probabilities.

  • The LSM input vector captures user context such as age, gender, occupation, and device type to predict content-request probabilities.
  • The output is a probability vector whose entries estimate each user’s probability of requesting each content.
  • The liquid model stores users’ dynamic contextual features over time for predicting content-request and mobility patterns.
  • After storing context and training the output function, the LSM predicts users’ content-request distributions from their context.

2) Resource allocation:

The resource-allocation procedure combines caching, user association, and licensed/unlicensed spectrum allocation. Search-based optimization derives resource allocations, while UAV agents use LSM learning to select association actions that maximize stable queues.

  • UAV agents observe association and network states, estimate stable-queue users, and use LSM outputs to choose resource-allocation actions.
  • The optimal cache contents for each UAV are the contents with the highest average request counts among its associated users.
  • The search-based algorithm finds an optimal resource allocation for a given user association, although the optimal allocation need not be unique.
  • Theorem 2 characterizes optimal licensed and unlicensed resource-allocation vectors for UAVs from candidate maximum-allocation vectors.

B. LSM Algorithm for content prediction and spectrum allocation

The complete LSM algorithm combines cloud-based content prediction with UAV-based learning for user association and spectrum allocation. It updates state and action information over time until each UAV’s output converges to an expected stable-queue value.

  • The cloud first predicts each user’s content-request distribution, after which UAVs use LSM learning with an ǫ-greedy mechanism to select user-association policies.
  • Given user association, Theorem 1 determines cached contents and the search-based algorithm determines spectrum allocation.
  • Each UAV records stable-queue users for alternative actions, and its resource-allocation output converges to the expected stable-queue value over other UAVs’ strategies.

IV. SIMULATION RESULTS

Simulations evaluate LSM-based prediction, caching, user association, and resource allocation against Q-learning and heuristic-search baselines. The proposed approach improves stable-queue users, data-rate distributions, prediction error, and convergence while exploiting caching and historical user information.

  • Up to 33.3% and 50.3% more users have stable queues than Q-learning with cache and without cache, respectively, with 5 UAVs.The gains are attributed to using historical user information for association and content-request prediction, which improves caching.
  • The LSM approach nearly matches heuristic search while finding optimal UAV resource allocation through a search-based algorithm.The search method adjusts its searching approach to users’ data-rate requirements and is reported to achieve up to a 28.5% gain in searches needed.
  • At 1.8 Mbps, the proposed approach improves the data-rate CDF by up to 27.56% and 48% over Q-learning with cache and without cache, respectively.Across the considered algorithms, almost 100% of users have data rates below 2 Mbps.
  • LSM prediction stays within 10% of real content-request probabilities and reduces mean squared error by 13.4% versus ESN.The prediction gap does not change request-probability rankings, allowing the cloud to identify optimal cached contents.
  • The proposed approach converges in 400 iterations, using 33.3% fewer iterations than Q-learning with cache.Stable-queue users increase until convergence, while LSM uses historical information to improve the learning process.

V. CONCLUSION

The paper develops an LSM-based framework for cache-enabled UAVs serving LTE-U users, jointly targeting stable queues through caching, prediction, and resource allocation. The proposed approach predicts content requests, determines cached contents, and enables autonomous spectrum allocation with limited network-state information.

  • V. CONCLUSION: The proposed LSM framework targets maximizing the number of users with stable queues in an LTE-U system using cache-enabled UAVs.
  • V. CONCLUSION: LSM predicts each user's content-request distribution so the cloud can determine UAV cached contents.
  • V. CONCLUSION: Each UAV autonomously selects its spectrum-allocation scheme using limited information about the network state.
  • V. CONCLUSION: Simulation results show significant performance gains for the proposed approach over the considered learning baselines.
  • V. CONCLUSION: LSM significantly improves convergence time compared with Q-learning.
Loading 1801.09339v1…