💻 computer science

DOA-EDS: Obstacle-Aware Edge Server Placement and Dynamic Task Scheduling for Industrial IoT

This paper proposes DOA-EDS, a two-stage framework combining an Obstacle-Aware Adaptive Large-Neighborhood Search for edge server placement and a feasibility-masked Proximal Policy Optimization for dynamic task scheduling, to significantly enhance network throughput and reduce latency in Industrial IoT environments plagued by severe signal blockages from metallic obstacles.

Original authors: Jingbo Ji, Fuyu Liu, Yong Liu, Xuejian Chi

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jingbo Ji, Fuyu Liu, Yong Liu, Xuejian Chi

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

Imagine you are running a massive, high-speed delivery service, but instead of trucks and roads, you are moving digital data. In the world of the "Industrial Internet of Things" (IIoT), sensors and machines are constantly shouting out information about their status. Usually, this data travels all the way to a giant, distant cloud server to be processed. But that's like sending a letter to the other side of the world just to ask for the time; it takes too long, and for things like oil rigs or construction sites, that delay can be dangerous. To fix this, engineers use "Edge Computing," which is like setting up small, local post offices right next to the workers. These local servers can process data instantly. However, there's a catch: these industrial sites are often filled with giant, solid metal obstacles—like massive storage tanks—that act like invisible walls, blocking the radio signals between the workers and the servers. If the signal hits a tank, the message is lost.

So, the big question becomes: How do you place these local servers so they can "see" the workers without hitting the metal walls, and how do you decide which server handles which job when the workers are moving around randomly? This is the puzzle tackled by a team of researchers who proposed a new system called DOA-EDS. They realized that simply guessing where to put the servers or using standard computer tricks doesn't work because the metal tanks create hard, physical blockages that break the math used by older methods. Instead, they built a two-step "smart planner" that first figures out the best safe spots to put the servers and then uses a learning computer brain to assign tasks in real-time.

The Problem: A Game of "Don't Hit the Tanks"

Imagine you are playing a game of tag in a warehouse filled with giant, impenetrable steel barrels. You have five "catchers" (the edge servers) who need to tag as many "runners" (the data tasks) as possible. The catchers can only tag a runner if they have a straight, unbroken line of sight. If a steel barrel is between them, the tag fails.

In the past, researchers tried to solve this by placing catchers on pre-existing shelves (like old cell towers) or by using simple math that assumed the air was empty. But in a real oil field, there are no shelves, and the air is full of tanks. If you place a server behind a tank, it's useless. If you place it too close, the signal bounces off the metal and fails. The researchers found that trying to solve this all at once—deciding where to stand and who to tag simultaneously—is incredibly difficult, so difficult that it's mathematically classified as "NP-hard," meaning it's a nightmare for computers to solve perfectly in a short time.

The Solution: A Two-Stage Teamwork Strategy

The authors of this paper, Jingbo Ji and their team, designed a clever two-stage framework called DOA-EDS to beat this challenge. Think of it as a coach and a referee working together.

Stage 1: The Coach (The Spatial Planner)
First, the system acts like a coach planning the game before it starts. It uses an algorithm called OA-ALNS (Obstacle-Aware Adaptive Large-Neighborhood Search). Imagine the coach looking at a map of the warehouse and simulating thousands of different ways to place the five catchers.

It knows the rules: "You cannot stand inside a barrel."
It knows the physics: "You cannot tag a runner if a barrel is in the way."
It tries a placement, sees if it's bad, and then "destroys" that bad idea and "repairs" it by moving the catchers to new spots.
It repeats this over and over, learning which spots give the best view of the runners while avoiding the tanks.

This stage happens offline (before the game starts) and finds the perfect, fixed coordinates for the servers. It doesn't guess; it systematically searches for the best "communication corridors" where signals can flow freely.

Stage 2: The Referee (The Dynamic Scheduler)
Once the servers are physically placed, the game begins. The runners (tasks) start appearing randomly. Now, the system switches to the second stage: a Deep Reinforcement Learning agent using an algorithm called PPO (Proximal Policy Optimization).

Think of this as a super-fast referee who watches the game in real-time.
The referee sees which server is busy, which is free, and where the runners are.
Crucially, the referee has a "feasibility mask." This is like a special pair of glasses that instantly tells the referee, "Hey, you can't assign this runner to Server A because a tank is blocking the path."
The referee learns from experience. If it assigns a task to a server that gets overloaded, it learns not to do that next time. It constantly adjusts to keep the game running smoothly, balancing the load so no server is overwhelmed while others sit idle.

What They Found: The Results

The team tested their system in a computer simulation of a 250-meter by 250-meter oil field filled with four giant storage tanks. They compared their two-stage team against other methods, including systems that tried to do everything at once (end-to-end learning) and simple "greedy" methods that just picked the nearest server.

The results were quite clear in their simulations:

Success Rate: When the system was under heavy load with 250 tasks, the DOA-EDS system successfully handled 95.2% of them. In contrast, the "end-to-end" learning systems (which tried to learn placement and scheduling simultaneously) only managed about 68% to 70%. They struggled because they couldn't figure out the hard geometry of the tanks.
Speed (Latency): The average time it took to process a task with DOA-EDS was 248.3 milliseconds. The other learning-based system (DQN-Joint) was much slower, taking 1,099.8 milliseconds. This means DOA-EDS was 77.4% faster than that specific competitor.
Cost: Because DOA-EDS placed the servers so efficiently, it needed fewer resources to get the job done. The "cost" score for their method was 5.25, while the DQN-Joint method scored 13.12, meaning the older method would require roughly 150% more infrastructure to do a worse job.

Why It Matters

The paper suggests that by separating the problem into "where to put the servers" and "how to schedule the tasks," you can solve a problem that was previously too messy for computers. The "Coach" handles the hard, physical rules of the metal tanks, and the "Referee" handles the fast, changing flow of data.

The authors emphasize that this is a simulation-based study. They didn't build a physical oil field with real servers; they built a digital twin to test their ideas. However, the results suggest that for industrial sites filled with massive metal obstacles, this two-step approach is significantly better than trying to use a single, all-in-one learning model. It proves that sometimes, the best way to solve a complex problem is to break it down: first, find the safe path, and then run the race.

Technical Summary: DOA-EDS

Problem Statement

The paper addresses the critical challenge of deploying edge servers and scheduling tasks within Industrial Internet of Things (IIoT) environments, specifically focusing on oil and gas field stations. These environments present unique constraints that render traditional edge computing solutions ineffective:

Severe Geometric Obstacles: Massive, densely packed metallic storage tanks create absolute Line-of-Sight (LoS) blockages. Unlike urban scenarios where signal attenuation is probabilistic, these tanks cause total transmission failure if the LoS path is intersected.
Continuous Deployment Space: Unlike urban settings where servers are placed on pre-existing infrastructure (e.g., macro base stations), oil field stations often lack such infrastructure. Servers must be deployed at arbitrary 2D coordinates within a continuous space, subject to strict non-penetrable obstacle constraints.
Dynamic Task Loads: Construction elements (workers, machinery) generate data stochastically, leading to rapid spatiotemporal fluctuations in task distribution. Static scheduling algorithms fail to balance loads under these conditions, causing resource exhaustion on some servers while others remain idle.

The core problem is formulated as a coupled, NP-hard optimization task: determining optimal continuous 2D deployment coordinates for edge servers while simultaneously managing dynamic task scheduling, all while adhering to strict, non-differentiable geometric LoS constraints.

Methodology: The DOA-EDS Framework

The authors propose DOA-EDS (Deep Reinforcement Learning-assisted Obstacle-Aware Edge Deployment and Scheduling), a two-stage framework that decouples the spatial and temporal aspects of the problem to manage complexity.

Stage 1: Obstacle-Aware Adaptive Large-Neighborhood Search (OA-ALNS)

This stage handles the offline spatial deployment problem.

Challenge: The deployment space is continuous, and constraints (servers cannot be inside tanks; LoS paths cannot intersect tanks) are non-differentiable, making gradient-based optimization unsuitable.
Approach: The algorithm discretizes the valid physical space into a high-resolution grid, excluding tank interiors. It employs an Adaptive Large-Neighborhood Search (ALNS) with specific operators tailored to the obstacle environment:
- Destroy Operators: Remove poorly positioned servers based on "obstacle-shadow" (high blocked demand), "load-pressure" (overloaded capacity), or "redundancy" (overlapping coverage).
- Repair Operators: Re-insert servers into locations that maximize visibility gain, improve load balance, or utilize regret-based insertion logic.
Mechanism: A reward-driven selection mechanism adaptively weights these operators, while a decreasing record-to-record threshold allows for diversification in early iterations and intensification around high-quality feasible solutions in later iterations.

Stage 2: Feasibility-Masked Proximal Policy Optimization (PPO)

This stage handles the online dynamic task scheduling problem.

Challenge: Once servers are deployed, tasks arrive randomly. The scheduler must make millisecond-level decisions to balance load and minimize latency.
Approach: The problem is modeled as a Markov Decision Process (MDP) solved via Deep Reinforcement Learning (DRL).
- State Space: Includes real-time server computing/storage loads, incoming task demands, and a pre-calculated binary validity matrix ( $V(t)$ ) that encodes the fixed LoS constraints derived from Stage 1.
- Action Space: The agent sequentially assigns tasks to servers or rejects them. Crucially, a feasibility mask is applied before action selection to strictly prohibit assignments that violate LoS constraints or exceed server capacity.
- Algorithm: The authors utilize Proximal Policy Optimization (PPO) for its stability in handling constrained categorical decisions. The reward function is the negative weighted sum of service latency and unscheduled task penalties.

Key Contributions

Realistic System Modeling: The paper constructs a system model that rigorously formulates absolute geometric LoS blockages caused by metallic storage tanks, addressing a blind spot in existing MEC literature which often relies on simplified distance-decay or probabilistic models.
NP-Hard Formulation: The authors formally define the coupled deployment and scheduling problem and prove its NP-hard nature by reduction to the planar Euclidean p-center problem.
Two-Stage Framework: The proposal of DOA-EDS, which separates the spatial planning (OA-ALNS) from temporal scheduling (DRL-PPO). This allows the system to handle non-differentiable geometric constraints offline while leveraging DRL for online adaptability.
Obstacle-Specific Operators: The design of specific destroy-and-repair operators within OA-ALNS that explicitly account for obstacle shadows and LoS visibility, rather than generic spatial heuristics.

Experimental Results

Extensive simulations were conducted in a 250m × 250m environment with 4 storage tanks and varying task loads (50–250 tasks). The framework was compared against end-to-end DRL baselines (PPO-Joint, DQN-Joint), a robust heuristic (READ), and ablation variants.

Network Throughput: Under maximum load (250 tasks), DOA-EDS achieved a 95.2% task success rate. In contrast, end-to-end DRL baselines dropped to approximately 68–70%, and the heuristic READ achieved 78%. The authors attribute the DRL baselines' failure to their inability to learn hard geometric constraints effectively.
Service Latency: DOA-EDS reduced average service latency by 77.4% compared to the DQN-Joint baseline (248.3 ms vs. 1099.8 ms) under heavy load.
Deployment Cost: The framework demonstrated superior cost-effectiveness, incurring a normalized deployment cost of 5.25, which was significantly lower than the DQN-Joint (13.12) and READ (higher than DOA-EDS) methods.
Ablation Study: Removing the OA-ALNS stage (random placement) caused throughput to drop to 72.4%, confirming the necessity of geometric optimization. Removing the DRL stage (greedy scheduling) caused latency to more than double (520.3 ms vs. 248.3 ms), confirming the necessity of intelligent load balancing.

Significance and Claims

The paper claims that DOA-EDS provides a robust solution for IIoT scenarios where physical obstacles fundamentally alter signal propagation. By decoupling the spatial planning from temporal scheduling, the framework successfully navigates the "non-differentiable" nature of geometric constraints that typically hinder standard learning-based approaches.

The authors emphasize that their approach is particularly significant for safety-critical industrial applications (like oil and gas) where signal reliability is paramount. They assert that the synergy between the geometry-aware OA-ALNS (which establishes valid "communication corridors") and the load-aware DRL agent (which maximizes resource utilization within those corridors) is the key to achieving low-latency, high-throughput edge computing in obstacle-rich environments. The work concludes by noting that future research will explore map-derived layouts, interference effects, and larger-scale testbed validations.