PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning Based Caching for Self-Driving Car in Multi-access Edge Computing

Anselme Ndikumana, Nguyen H. Tran, Choong Seon Hong

arXiv:1810.01548v2cs.NI

TL;DR

Distant data-center retrieval can delay infotainment delivery to self-driving cars, while cars also need to select content suited to passengers. The paper combines area-specific MLP predictions, CNN-based passenger features, MEC caching, and delay-minimizing optimization. Simulations report 61% less backhaul traffic and 97.82% accuracy for predicting content probabilities across six RSU areas.

Abstract

from arXiv · show

Once self-driving car becomes a reality and passengers are no longer worry about it, they will need to find new ways of entertainment. However, retrieving entertainment contents at the Data Center (DC) can hinder content delivery service due to high delay of car-to-DC communication. To address these challenges, we propose a deep learning based caching for self-driving car, by using Deep Learning approaches deployed on the Multi-access Edge Computing (MEC) structure. First, at DC, Multi-Layer Perceptron (MLP) is used to predict the probabilities of contents to be requested in specific areas. To reduce the car-DC delay, MLP outputs are logged into MEC servers attached to roadside units. Second, in order to cache entertainment contents stylized for car passengers' features such as age and gender, Convolutional Neural Network (CNN) is used to predict age and gender of passengers. Third, each car requests MLP output from MEC server and compares its CNN and MLP outputs by using k-means and binary classification. Through this, the self-driving car can identify the contents need to be downloaded from the MEC server and cached. Finally, we formulate deep learning based caching in the self-driving car that enhances entertainment services as an optimization problem whose goal is to minimize content downloading delay. To solve the formulated problem, a Block Successive Majorization-Minimization (BS-MM) technique is applied. The simulation results show that the accuracy of our prediction for the contents need to be cached in the areas of the self-driving car is achieved at 98.04% and our approach can minimize delay.

I. Introduction

Self-driving cars create a need for personalized infotainment, but retrieving content from distant data centers can cause high delay. The paper proposes feature-aware caching using deep learning, MEC, and models of communication, caching, and computation.

A. Background and Motivations

B. Challenges for infotainment Caching

  • Passenger features such as age and gender shape content preferences, motivating feature-dependent caching decisions.The paper frames self-driving cars as needing to determine content appropriate for passengers themselves.Source: B. Challenges for infotainment Caching • · p. 2

C. Related Works

  • Prior work addressed vehicle infotainment analysis without caching, cloud-based caching with potential retrieval delay, and caching without occupant-based selection,,.The paper distinguishes its focus from autonomous-car caches for sensor and driving-decision data.Source: C. Related Works · p. 2

D. Contributions

  • The proposed framework combines area-specific MLP content-demand predictions at MEC servers with CNN passenger-feature predictions and classification to select car-cache content.The CNN application uses facial images to infer passenger features; the MLP output is stored at RSUs for nearby cars.Source: D. Contributions · pp. 2–3
  • The paper adds en-route RSU selection and computation models that let RSUs and cars retrieve content and process cached items into requested formats or qualities.The models account for available communication, caching, and computation resources.Source: D. Contributions · pp. 2–3
  • The authors formulate delay-minimizing caching as a non-convex optimization problem, use a convex surrogate with BS-MM, and identify feature- and resource-aware car caching as their novelty.BS-MM decomposes the problem into smaller subproblems that can be solved separately.Source: D. Contributions · p. 3

II. System model

The system places trained deep-learning models and predicted content demand near cars at RSUs, while cars use passenger features to choose cached infotainment. The section describes area-based MLP prediction, CNN feature inference, and processing cached content into requested formats.

II. System model

  • The DC trains and tests CNN and MLP models, while RSU MEC servers store model outputs and cache area-specific content for nearby cars.Each car can retrieve the models and MLP output from an RSU and has its own caching and computation resources.Source: §II · pp. 3–4

III. Deep Learning Based Caching in Self-Driving Cars

  • RSUs and cars can compute cached infotainment into demanded formats, such as converting cached MP4 content to H.264 instead of retrieving the same content again.The system model also covers communication, caching, and computation resources for retrieving and serving content.Source: §II · p. 4, §III · p. 4

A. Deep Learning and Recommendation Model

  • The MLP takes content names, ratings, viewer age, gender, and location as inputs and predicts content-request probabilities for geographic RSU areas.Its output layer has one neuron per area, and the model uses cross-entropy for this area-classification task.Source: 1) Multi-Layer Perceptron (MLP) · pp. 4–5
  • The paper applies an existing CNN to facial images to predict passengers’ age, emotion, and gender for caching decisions rather than proposing a new CNN.Cars compare CNN predictions with MLP outputs using k-means and binary classification to identify appropriate content.Source: §II · p. 4, 2) Convolutional Neural Network (CNN) · p. 5

2) Convolutional Neural Network (CNN):

The CNN classifies passenger facial images by age, gender, and emotion; the recommendation model combines those features with area-based content predictions to select content to cache.

2) Convolutional Neural Network (CNN):

  • The CNN applies convolution, ReLU, and max-pooling layers before fully connected layers classify facial images by age, gender, and emotion.Its three outputs cover two gender classes, 101 age classes, and eight emotion classes; softmax gives class probabilities.Source: 2) Convolutional Neural Network (CNN) · p. 5

3) Recommendation Model:

  • For each passenger, the car predicts age, gender, and emotion from a facial image, then assigns the passenger to a matching feature-based cluster.Age and emotion grouping use k-means; gender grouping uses binary classification.Source: 3) Recommendation Model · p. 6
  • The car selects high-probability contents for the passenger’s age, emotion, and gender clusters, then downloads and caches them via an RSU.The model describes selecting top predicted contents within the assigned clusters.Source: 3) Recommendation Model · p. 6
  • Cars and MEC servers are assumed to store the CNN model and MLP output during off-peak hours, leaving recommended-content downloading, caching, and computing as the focus.Source: 3) Recommendation Model · p. 6

B. Communication Model for Retrieving Contents

The communication model plans which en-route RSUs can supply recommended contents and represents the transmission delays for delivery to cars and their passengers.

B. Communication Model for Retrieving Contents

  • The model accounts for downloading area-predicted contents from the DC to MEC servers and recommended contents from MEC servers to cars over RSU wireless links.The RSU-to-car capacity is modeled using bandwidth, channel gain, and RSU transmission power.Source: B. Communication Model for Retrieving… · pp. 6–7
  • The car can select available en-route RSUs before departure to reduce transmission-delay variation and handoff, using route information returned by ANDSF.The request to ANDSF includes the car’s location, speed, and direction; its feedback provides coordinates and coverage of RSUs in the car’s direction.Source: B. Communication Model for Retrieving… · pp. 6–7
  • Passenger delivery over the car’s WiFi uses an instantaneous rate that accounts for throughput efficiency, connected-passenger count, and channel utilization.The utilization function represents the impact of contention on WiFi throughput.Source: B. Communication Model for Retrieving… · p. 7

C. Caching Model for Retrieved Contents

The caching model represents content placement and cache hits at both cars and MEC servers, subject to limited storage.

C. Caching Model for Retrieved Contents

D. Computation Model for Cached Contents

The computation model accounts for transcoding cached content into the requested format at either a car or MEC server, subject to computational-resource limits.

D. Computation Model for Cached Contents

  • If a cached item is in a different format, it can be transcoded to the requested format, but the model assumes finite computation resources.When a car lacks sufficient computation resource, it forwards the request to the MEC server.Source: D. Computation Model for Cached Contents · p. 8
  • The car allocates computation resources for format conversion using weighted proportional allocation, with workload measured in CPU cycles per bit.Source: D. Computation Model for Cached Contents · p. 8
  • Constraints require conversion to occur at exactly one location and ensure the car or MEC server returns only one format of the requested content.If resource and caching constraints cannot be met, the MEC server submits the request to the DC.Source: D. Computation Model for Cached Contents · pp. 8–9

IV. Problem Formulation and Solution

The section introduces an optimization problem for minimizing infotainment-download delay and then presents a solution.

  • The section formulates an infotainment-download delay minimization problem and presents its solution in the following subsections.Source: §IV · p. 9

A. Problem Formulation

The formulation links the car’s learning, recommendation, communication, caching, and computation models in an optimization problem that minimizes infotainment retrieval delay. Retrieval can proceed from the car to an RSU and, if needed, the DC, subject to resource and format constraints.

A. Problem Formulation

  • The optimization links learning, recommendation, communication, caching, and computation models to minimize infotainment retrieval delay.Source: A. Problem Formulation · p. 9
  • Requested content is retrieved from the car when available, otherwise from an RSU, and from the DC as a last resort.Source: A. Problem Formulation · p. 9
  • Caching and computation allocations must fit available resources, while only one content format and one conversion location are allowed.Caching decisions are based on CNN output and passenger features such as age, emotion, and gender.Source: A. Problem Formulation · p. 9
  • The formulated optimization is non-convex, so the authors propose a proximal convex surrogate and apply BS-MM to solve it.Source: A. Problem Formulation · p. 9

B. Proposed Solution: Distributed Algorithm for Deep Learning Based Caching

The proposed distributed algorithm uses BS-MM to solve a proximal convex surrogate through iterative block updates, then rounds relaxed variables and checks resource feasibility. The authors state that it converges to a coordinate-wise stationary point under the stated condition.

B. Proposed Solution: Distributed Algorithm for Deep Learning Based Caching

V. Simulation Results and Analysis

The performance evaluation uses Google Maps Services for mobility analysis, Keras with TensorFlow for deep-learning simulation, and pandas for data analysis.

V. Simulation Results and Analysis

  • The evaluation uses Google Maps Services for mobility analysis, Keras with TensorFlow for deep-learning simulation, and pandas for data analysis.Source: §V · p. 11

A. Simulation Setup

The simulation combines Movie-Lens data with synthetic content and passenger attributes, then evaluates regional demand prediction using configured RSU, vehicle, and network resources.

A. Simulation Setup

  • The Movie-Lens dataset supplies movie genres and user age, gender, ratings, and ZIP codes; movie sizes and formats are randomly generated for the caching model.Generated movie sizes range from 317 to 750 Mb, and ZIP codes are converted to coordinates to locate users.Source: A. Simulation Setup · pp. 11–12, Figure 5 · p. 12
  • Six RSU areas are selected from movie-watching counts, rankings, and user locations, and a two-hidden-layer MLP predicts regional content-request probabilities.The model uses 32 neurons per layer except for its six-neuron output, with a 60%/40% train-test split, learning rate 0.002, and batch size 32.Source: A. Simulation Setup · p. 12
  • The route runs from RSU 1 to RSU 6; each RSU has 60–70 Mbps wired backhaul, 10 MHz bandwidth, 3.6 GHz CPU, and 100–110 TB cache.RSU locations are updated to create realistic distances, and route speed and RSU connectivity are determined using distance and duration.Source: A. Simulation Setup · p. 12
  • The car uses randomly generated features for 37 passengers, with eight age and eight emotion clusters, and content demand follows a Zipf distribution.The car is configured with 160 MHz WiFi, 3.6 GHz computation capacity, and 100 TB cache.Source: A. Simulation Setup · pp. 12–13

B. Evaluation Results

Deep3C combines regional MLP demand prediction with passenger-feature-based in-car caching, then evaluates prediction, delay optimization, and computational resource use. The reported results include 61% of infotainment demands served by caching and 97.82% prediction accuracy.

B. Evaluation Results

  • 97.82% accuracy is achieved for predicting content-request probabilities across six RSU areas; RSUs cache high-rated, high-probability movies until storage fills.The six RSU areas are selected using k-means clustering of ratings and location information.Source: B. Evaluation Results · p. 13
  • The car combines downloaded MLP output with passenger age, emotion, and gender clusters to select highly rated, likely-to-be-requested movies for its passengers.The evaluation uses eight age clusters, eight emotion clusters, and two gender clusters; CNN can predict the passenger features from camera images.Source: B. Evaluation Results · pp. 13–14
  • The delay-minimization solution converges to a coordinate-wise stationary point under Cyclic, Gauss-Southwell, and Randomized selection, with high latency during the initial content download.Once cached, similar passenger requests can be served from car storage.Source: B. Evaluation Results · p. 14, Figure 13 · p. 14
  • Cyclic selection uses more computational resource than Gauss-Southwell and Randomized selection in the OBU throughput comparison.The comparison reports computational throughput as a CDF in instructions per second.Source: B. Evaluation Results · p. 14, Figure 14 · p. 14

VI. Conclusion

  • 61% of whole infotainment demands are served by caching at RSUs and self-driving cars, and prediction of contents to cache reaches 97.82% accuracy.The proposed optimization targets content-downloading delay.Source: §VI · p. 15
Loading 1810.01548v2…