[Paper Review] Task Replication for Vehicular Edge Computing: A Combinatorial Multi-Armed Bandit based Approach

Yuxuan Sun, Jinhui Song+3 more|arXiv (Cornell University)|Jul 16, 2018

IoT and Edge/Fog Computing14 references3 citations

TL;DR

This paper proposes a learning-based task replication algorithm (LTRA) using combinatorial multi-armed bandit (CMAB) theory to minimize average offloading delay in highly dynamic vehicular edge computing (VEC) systems. By enabling task replication across multiple service vehicles and learning channel and load conditions in real time, LTRA improves service reliability, achieving a 98% task completion ratio within 0.6s—up from 80% with single offloading.

ABSTRACT

In vehicular edge computing (VEC) system, some vehicles with surplus computing resources can provide computation task offloading opportunities for other vehicles or pedestrians. However, vehicular network is highly dynamic, with fast varying channel states and computation loads. These dynamics are difficult to model or to predict, but they have major impact on the quality of service (QoS) of task offloading, including delay performance and service reliability. Meanwhile, the computing resources in VEC are often redundant due to the high density of vehicles. To improve the QoS of VEC and exploit the abundant computing resources on vehicles, we propose a learning-based task replication algorithm (LTRA) based on combinatorial multi-armed bandit (CMAB) theory, in order to minimize the average offloading delay. LTRA enables multiple vehicles to process the replicas of the same task simultaneously, and vehicles that require computing services can learn the delay performance of other vehicles while offloading tasks. We take the occurrence time of vehicles into consideration, and redesign the utility function of existing CMAB algorithm, so that LTRA can adapt to the time varying network topology of VEC. We use a realistic highway scenario to evaluate the delay performance and service reliability of LTRA through simulations, and show that compared with single task offloading, LTRA can improve the task completion ratio with deadline 0.6s from 80% to 98%.

Motivation & Objective

• Address the challenge of highly dynamic vehicular networks where channel states and computation loads vary rapidly, making traditional modeling or prediction infeasible.
• Improve quality of service (QoS) in VEC by leveraging redundant computing resources from dense vehicle populations.
• Design a distributed, learning-based task offloading mechanism that enables task vehicles to adaptively select service vehicles without global state knowledge.
• Minimize average offloading delay while maintaining high service reliability under time-varying network conditions.
• Optimize the number of task replicas and discretization level to balance performance and computational cost in real-world scenarios.

Proposed method

• Proposes a distributed task replication framework where each task vehicle (TaV) offloads replicas to multiple candidate service vehicles (SeVs) simultaneously.
• Applies combinatorial multi-armed bandit (CMAB) theory to model the selection of SeV groups, treating each group as a 'combinatorial arm' with a delay-based reward.
• Redesigns the utility function to incorporate vehicle occurrence times, enabling adaptation to time-varying VEC topology.
• Uses a confidence-bound approach to estimate cumulative distribution functions (CDFs) of delay, balancing exploration and exploitation.
• Transforms the delay minimization problem into a reward maximization problem via normalization and monotonic reward transformation.
• Employs a discretized delay estimation method with tunable level $ l $, optimizing the trade-off between runtime and estimation accuracy.

Experimental results

Research questions

• RQ1 How can task offloading delay be minimized in a highly dynamic VEC environment with limited global state information?
• RQ2 What is the optimal number of task replicas to balance delay reduction and resource utilization under varying traffic densities?
• RQ3 How does the discretization level of delay distributions affect the performance and runtime of the learning algorithm?
• RQ4 Can a learning-based approach outperform single offloading in terms of both average delay and service reliability?
• RQ5 How does the algorithm adapt to time-varying vehicle topology and transient channel conditions in real highway scenarios?

Key findings

• LTRA reduces average offloading delay significantly compared to single offloading, approaching the performance of a genie-aided optimal policy.
• With a deadline of 0.6 seconds, LTRA achieves a 98% task completion ratio using only two replicas, compared to 80% with single offloading.
• When the task vehicle to service vehicle density ratio exceeds 0.3, using fewer replicas (e.g., 2) yields better delay performance than more replicas (e.g., 4).
• A discretization level $ l $ of 40–50 provides the best trade-off between delay estimation accuracy and algorithm runtime.
• The algorithm’s learning regret is theoretically bounded, confirming its convergence and stability under the CMAB framework.
• LTRA effectively adapts to dynamic vehicular topologies by incorporating vehicle occurrence times into the utility function, improving real-time adaptability.