💻 computer science

Pressure Guided Hierarchical Offloading and Heterogeneous Resource Selection for UAV-assisted Vehicular Edge Computing Networks: An Perspective of Future Task Prediction

This paper proposes Pressure Guided Hierarchical Offloading (PGHO), a two-tier decision-making framework for UAV-assisted vehicular edge computing that integrates future task pressure predictions with current resource conditions to significantly reduce average delay and improve task success rates compared to existing multi-agent reinforcement learning baselines.

Original authors: Chao He, Dongfeng Fu, Xin Xie, Jinkui Zhang, ShuXu Yue, Zheng Zhang

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chao He, Dongfeng Fu, Xin Xie, Jinkui Zhang, ShuXu Yue, Zheng Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the bustling world of modern transportation, cars are no longer just machines that move people from one place to another; they are becoming rolling computers. As vehicles gain the ability to recognize objects, understand complex environments, and make split-second safety decisions, they generate a massive amount of data that needs to be processed instantly. However, the computers inside these cars have limits. They cannot handle every heavy calculation without slowing down or running out of battery. To solve this, engineers have developed a system called vehicular edge computing. Instead of trying to do everything inside the car, the vehicle sends its heavy tasks to nearby helpers. These helpers can be other cars, fixed stations on the side of the road, or even flying drones. The goal is to find the best helper for each job so that the car gets its answer quickly and safely.

The challenge lies in the chaos of the road. Traffic changes, weather shifts, and the needs of the drivers fluctuate wildly. A helper that looks perfect right now might be overwhelmed by a sudden rush of urgent tasks just a moment later. If a car sends a task to a helper that is about to become too busy, the task gets stuck, the car waits, and safety could be compromised. Researchers have long tried to predict these changes, but most systems only look at what is happening right now. They decide where to send a task based on the current queue length or signal strength, missing the fact that a burst of new, urgent jobs might be arriving in the next few seconds. This gap between the present moment and the immediate future is where delays and failures often begin.

A team of researchers has proposed a new approach to this problem, designed specifically for networks that include flying drones to assist ground vehicles. They call their method "Pressure Guided Hierarchical Offloading." Instead of making a single, complicated decision all at once, their system breaks the problem into two clear steps. First, it looks ahead. It gathers information about what kind of tasks are likely to arrive, how urgent they will be, and how much work they will require in the very near future. It combines all these predictions into a single measure of "future task pressure." This allows the system to see not just the current load, but the weight of what is coming down the pipeline.

Once this future pressure is understood, the system makes a high-level choice: which helper should take the job? This could be the car itself, a nearby vehicle, a roadside station, or a drone. The researchers found that by considering the predicted pressure before making this choice, the system can avoid picking a helper that looks free now but will be swamped in seconds. After the helper is chosen, the system moves to the second step. It adjusts the drone's position and decides exactly how much of the task to send versus how much to keep local. Because the helper has already been selected, the system can focus entirely on optimizing the connection to that specific helper, rather than trying to juggle all possible options at once.

The researchers tested this method in a simulated environment that mimicked three very different driving conditions: normal traffic, heavy congestion, and situations where tasks had extremely tight deadlines. In the normal conditions, the new method reduced the average time it took to complete a task to about 0.398 seconds, while successfully finishing nearly 99.7 percent of all tasks. When the traffic became heavy and the competition for computing power increased, the system maintained its high success rate, completing 99.5 percent of tasks with an average delay of just 0.483 seconds. This is a significant improvement over other advanced methods, which often saw their success rates drop or their delays spike under the same pressure. In the most difficult scenario, where tasks had to be finished almost immediately, the system achieved a success rate of 99.85 percent with an average delay of 0.409 seconds.

The study also revealed why this two-step approach works so well. When the researchers removed the ability to predict future tasks, the average delay increased significantly, showing that looking ahead is crucial for timing. When they removed the two-step structure and tried to make all decisions at once, the delays grew even larger, proving that separating the choice of helper from the adjustment of the connection makes the system more efficient. Perhaps most importantly, the system included a safety check that prevented it from choosing helpers that were already too far away or had poor connections. This simple filter was vital when deadlines were tight, preventing the system from wasting time on impossible choices.

The researchers did not claim that this system is perfect or that it works in every real-world situation without testing. Their results come from computer simulations using specific models of how drones and cars communicate. They noted that the system uses more energy than a simple, greedy approach that just picks the closest helper, but this extra energy is the price paid for the much higher reliability and speed. The study suggests that by combining a look into the immediate future with a structured way of making decisions, we can build vehicle networks that are not only faster but also far more dependable when the road gets crowded and the pressure mounts.

Technical Summary: Pressure Guided Hierarchical Offloading and Heterogeneous Resource Selection for UAV-assisted Vehicular Edge Computing Networks

Problem Statement
Vehicular Edge Computing (VEC) networks, which integrate local vehicles, cooperative vehicles, Unmanned Aerial Vehicles (UAVs), and Roadside Units (RSUs), face significant challenges in dynamic environments. Existing approaches often fail to fully account for "future task pressure"—the anticipated evolution of workload, task arrivals, urgency, and deadlines in subsequent time slots. Furthermore, the heterogeneity of resources (aerial vs. ground) and the coupling of discrete decisions (selecting an execution node) with continuous controls (UAV trajectory and offloading ratios) complicate real-time optimization. Traditional static optimization is impractical due to high-dimensional states and real-time constraints, while standard Deep Reinforcement Learning (DRL) approaches often struggle to coordinate these mixed action types effectively or utilize predictive information to preemptively manage queue buildup and resource contention.

Methodology: The PGHO Framework
The paper proposes Pressure Guided Hierarchical Offloading (PGHO), a framework designed to integrate future task prediction with a hierarchical decision structure for UAV-assisted VEC.

Future Task Pressure Representation:
Instead of relying solely on current queue states, PGHO constructs a "Future Task Pressure" metric before resource selection. This involves predicting four key attributes over a short horizon ( $H=4$ slots): computational load, task urgency, arrival intensity, and deadline changes. These are processed using a multi-scale predictor (utilizing three time windows: short, medium, and long) to generate an exponentially weighted moving average (EWMA) and trend estimates. These predictions are aggregated into a compact Future Task Flow Pressure Index (FTPI) and a pressure vector, which are fed into the decision state to anticipate queue pressure and deadline risks before a node is chosen.
Hierarchical Decision Structure:
PGHO decomposes the joint decision problem into two levels based on dependency:
- Upper Policy (Discrete): Uses a Double Deep Q-Network (DDQN) to select the execution node (local vehicle, cooperative vehicle, UAV, or RSU). This decision is made using current resource conditions and the predicted future task pressure. Crucially, an action masking mechanism filters out physically infeasible nodes (based on coverage, link reliability, and transmission time) before the policy ranks candidates.
- Lower Policy (Continuous): Once a node is selected, a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) controller operates. This level manages UAV displacement and the offloading ratio ( $\rho$ ) specifically under the link, queue, and resource constraints of the already selected node. This hierarchy ensures that continuous control variables are generated within the feasible context established by the discrete choice.
Policy Selection and Training:
To address the instability of DRL training where the final policy may not be the best, PGHO employs a validation-based selection strategy. During training, policies are periodically evaluated on a fixed validation task flow. Checkpoints are retained based on a composite score prioritizing reliability (success ratio, failure types) and delay metrics (average, P95, P99). If performance degrades persistently, the system restores a previously superior checkpoint. This ensures the deployed policy is the one with the best observed validation performance, not merely the final training update.

Key Contributions
The paper identifies three primary contributions:

Future Task Pressure Representation: A novel method that jointly summarizes predicted workload, arrivals, urgency, and deadline changes into a unified pressure metric. This allows the upper-level decision to anticipate future demand evolution before selecting heterogeneous resources.
Hierarchical Decision Structure: A dependency-driven architecture where the upper policy selects the execution node, and the lower policy controls UAV movement and offloading ratios based on the conditions of that specific node. This separates discrete node selection from continuous control, reducing the complexity of the action space.
Validation-Based Policy Selection: A training procedure that retains policies with superior validation performance on a fixed task flow, utilizing checkpoint storage, rollback mechanisms, and candidate refinement to ensure robustness against training fluctuations.

Experimental Results
The method was evaluated across three scenarios: normal load, high load, and tight deadlines, using random seeds 42, 43, and 44. PGHO was compared against TF Greedy, MADDPG, Soft Actor-Critic (SAC), and Hybrid Multi-Agent Proximal Policy Optimization (H MAPPO).

Performance: Across all scenarios, PGHO achieved an average delay of 0.4301 s and a task success ratio of 99.705%.
Comparisons:
- Compared to MADDPG, SAC, and H MAPPO, PGHO reduced average delay by 21.1%, 12.1%, and 7.4%, respectively, while simultaneously increasing the task success ratio.
- Compared to the competitive TF Greedy baseline, PGHO maintained a slight advantage in average delay (approx. 1.83% better) while improving the task success ratio by 2.62 percentage points.
Component Analysis:
- Hierarchical Structure: Removing the hierarchy (PGHO no HD) increased average delay by ~32.3%, demonstrating that separating node selection from continuous control is critical for latency reduction, especially under high load.
- Future Task Information: Removing prediction (PGHO no PE) increased average delay by ~33.6%, confirming that anticipating future queue pressure significantly improves timing and destination selection.
- Action Masking: Removing masking (PGHO no AM) caused a sharp drop in task success ratio under tight deadlines (from 99.85% to 93.41%), highlighting its role in preventing infeasible assignments.
- Validation Selection: The full selection process (vs. PGHO no SI) improved average delay by ~32.0%, proving that retaining the best validation checkpoint is superior to using the final training state.

Significance and Claims
The paper claims that PGHO effectively addresses the trade-off between low latency and high task completion in dynamic, heterogeneous VEC networks. By integrating future task pressure into the resource selection phase and decoupling discrete and continuous controls, the method achieves high reliability (near 100% success) without sacrificing mean delay performance. The authors note that while PGHO incurs a higher normalized energy cost (approx. 24.2% higher than TF Greedy) due to more active resource utilization, this is a necessary trade-off for achieving the observed reliability and low tail latency. The study concludes that the proposed framework maintains low delay and high task completion across varying task pressure conditions, offering a robust solution for future intelligent transportation systems. The authors acknowledge limitations regarding the use of simulated traces, a single wireless model, and fixed network sizes, suggesting future work should incorporate measured link conditions and real-world vehicle trajectories.

Technical Summary: Pressure Guided Hierarchical Offloading and Heterogeneous Resource Selection for UAV-assisted Vehicular Edge Computing Networks

More like this