🔢 mathematics

Joint Task Orchestration and Resource Optimization for SC3 Closed Loop in 6G Networks

This paper proposes a learning-optimization-integrated actor-critic (LOAC) framework to solve the joint sensor-actuator pairing and resource allocation problem in 6G SC3 closed loops, achieving near-optimal control cost reduction with low computational complexity.

Original authors: Xinran Fang, Wei Feng, Yanmin Wang, Yunfei Chen, Baoquan Ren, Ning Ge, Shi Jin

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinran Fang, Wei Feng, Yanmin Wang, Yunfei Chen, Baoquan Ren, Ning Ge, Shi Jin

Original paper licensed under CC BY 4.0 (http://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 a chaotic disaster zone, like a collapsed building after an earthquake. Humans can't go in safely, so we send in a swarm of robots. But here's the catch: a single robot is like a person with one eye and one hand. It can see a bit, but it can't do much. To save lives, these robots need to work together as a single, super-efficient team.

This paper is about how to organize that team using 6G technology (the super-fast internet of the future). The authors propose a system where a "brain" (called an Edge Information Hub or EIH) flies overhead, connecting sensors (the eyes) and robots (the hands) to create a perfect loop of action.

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Blind Date" of Robots

In a normal network, you just connect one device to the internet. But in this disaster scenario, the system has to do something much harder: match a specific sensor to a specific robot.

The Challenge: You have 20 cameras (sensors) scattered around the rubble and only 4 rescue robots. Which camera should talk to which robot?
The Trap: If you pick the wrong pair, the robot might be too far from the camera, or the camera might be too "heavy" (sending too much data) for the robot's computer to handle.
The Coupling: It's not just about picking the pair; it's also about how much "fuel" (bandwidth and computing power) you give them. If you give a robot a great camera but no computing power, it can't process the image. If you give it a great computer but a bad connection, it gets no data.

The goal is to minimize the "LQR Cost." Think of this as the "Stress Score" of the system. A high stress score means the robots are wobbling, crashing, or failing to act fast enough. A low score means they are moving smoothly and saving lives.

2. The Solution: The "LOAC" Framework

The authors created a smart system called LOAC (Learning-Optimization-Integrated Actor-Critic). Imagine this as a Chess Coach and a Grandmaster working together.

The Actor (The Chess Coach):
- This is a computer brain (a Deep Neural Network) that looks at the situation.
- It says, "Okay, I think Camera A should pair with Robot 1, and Camera B with Robot 2."
- It makes a quick guess based on what it has learned. It's fast, but it might not be perfect.
The Critic (The Grandmaster):
- This is a mathematical optimizer. It takes the Coach's guess and runs a rigorous calculation.
- It says, "If you pair them that way, here is exactly how much bandwidth and computing power you need to make it work perfectly."
- It calculates the final "Stress Score" (LQR cost).
The Feedback Loop:
- The Critic tells the Coach: "That pairing was okay, but if you swapped Robot 1 and Robot 2, the Stress Score would drop by 20%."
- The Coach learns from this feedback. Over time, the Coach gets so good at guessing that it almost never needs the Grandmaster to correct it. It learns to make near-perfect decisions instantly.

3. Why This is Better Than Old Methods

The paper compares their method to three other ways people try to solve this:

Exhaustive Search (The "Try Everything" Method):
- Imagine trying every single possible combination of cameras and robots.
- Verdict: It finds the absolute best answer, but it takes so long (like trying to find a needle in a haystack by checking every straw) that the disaster is over before you finish. It's too slow for real-time rescue.
Decoupled Optimization (The "Step-by-Step" Method):
- First, pick the pairs. Then, figure out the resources.
- Verdict: This is like buying a suit without trying it on, then trying to tailor it later. You often end up with a bad fit because the two steps weren't planned together.
Communication-First (The "Fastest Pipe" Method):
- Just pick the pairs with the strongest internet signal.
- Verdict: This ignores the robot's brain. You might give a robot a super-fast camera, but if the robot can't process the video, the speed doesn't matter.

4. The Result: A Self-Adjusting Team

The LOAC system is special because it treats the Sensor + Connection + Computer + Robot as one single unit.

Dynamic Adaptation: If a robot moves or a building collapses (changing the environment), the system instantly re-evaluates. It doesn't just stick to a plan; it adapts on the fly.
Efficiency: It finds a solution that is almost as good as the "Try Everything" method but does it in a fraction of a second.

The Big Picture

In the past, network engineers focused on making the "pipes" (internet) faster. This paper says, "That's not enough." In a rescue mission, the goal isn't just to move data; it's to move the world (save lives).

By using this "Coach and Grandmaster" AI system, we can orchestrate a swarm of robots to act like a single, highly intelligent organism, ensuring that every sensor is paired with the right robot, and every robot gets exactly the resources it needs to perform its task perfectly. It turns a chaotic disaster zone into a synchronized, life-saving dance.

1. Problem Statement

The paper addresses the challenge of optimizing Sensing-Communication-Computing-Control (SC3) closed loops in 6G-enabled autonomous operation systems, specifically in hazardous environments (e.g., disaster rescue) where human intervention is impossible.

Core Challenge: In these scenarios, an Edge Information Hub (EIH) must coordinate distributed sensors and actuators to form autonomous "robots" (SC3 loops). The system faces a Joint Sensor-Actuator Pairing and Resource Allocation problem.
The Coupling Issue:
- Discrete Decisions: The EIH must select which sensor pairs with which actuator (a combinatorial problem).
- Continuous Resources: Once paired, the system must allocate continuous resources: uplink/downlink bandwidth, transmit power, and CPU computing frequency.
- Interdependence: Sensor selection dictates uplink communication and computing loads, while actuator selection dictates downlink requirements and control dynamics. These factors are tightly coupled; a poor pairing cannot be fixed by resource allocation, and vice versa.
Optimization Goal: Minimize the sum Linear Quadratic Regulator (LQR) cost across all $K$ SC3 loops. The LQR cost quantifies control performance (state deviation and control effort); a lower cost implies better stability and efficiency.
Mathematical Formulation: The problem is formulated as a Mixed-Integer Nonlinear Programming (MINLP) problem, which is NP-hard and computationally prohibitive for real-time decision-making using traditional exhaustive search methods.

2. Methodology: The LOAC Framework

To solve the complex MINLP problem efficiently, the authors propose a Learning-Optimization-Integrated Actor-Critic (LOAC) framework. This hybrid approach combines Deep Neural Networks (DNN) for discrete decisions with convex optimization for continuous variables.

A. System Modeling

SC3 Closed Loop: Modeled as a human reflex arc: Sensor (eyes) $\to$ EIH (brain) $\to$ Actuator (hands).
Metric (CNER): The authors introduce the Closed-Loop Negentropy Rate (CNER), a loop-level metric that unifies the performance of sensing, communication, computing, and control. It represents the entropy reduction of the controlled system per cycle.
Objective: Minimize LQR cost subject to CNER constraints, ensuring system stability ( $D^{SC3}_k > \log_2 |\det A_k|$ ).

B. The LOAC Architecture

The framework consists of two interacting modules:

Actor Module (DNN-based):
- Input: Encoded traits of sensors (location, channel gain, data volume, computational intensity) and actuators (location, channel gain, system dynamics).
- Function: A Deep Neural Network generates a raw pairing probability matrix ( $Y_{raw}$ ).
- Output: A valid, one-to-one sensor-actuator pairing solution ( $A$ ) is generated via a sequential selection algorithm (Algorithm 1) that respects sensing range constraints and randomizes actuator order to avoid bias.
Critic Module (Optimization-based):
- Function: Given a specific pairing decision from the Actor, the Critic solves the resulting continuous resource allocation problem (minimizing LQR cost by optimizing bandwidth, power, and CPU frequency).
- Convexity: The authors prove that once the pairing is fixed, the resource allocation problem becomes a convex optimization problem (Theorem 1), allowing for efficient, globally optimal solutions using standard tools (e.g., interior-point methods).
- Feedback: The Critic calculates the resulting LQR cost ( $L^*_i$ ) for the pairing. This cost serves as a precise "reward" or feedback signal.
Interaction & Training:
- The Critic selects the best pairing solution from a batch of candidates generated by the Actor.
- This optimal pairing is stored in a replay buffer.
- The Actor's DNN is updated via supervised learning (minimizing cross-entropy loss) using the Critic's optimal decisions as ground truth.
- This iterative process allows the DNN to learn the complex mapping between system states and optimal pairings without needing to solve the full MINLP during inference.

3. Key Contributions

System-Level SC3 Optimization: The paper proposes a novel joint optimization scheme for sensor-actuator pairing and resource allocation, moving beyond traditional approaches that assume fixed loop configurations.
Introduction of CNER: A new metric (CNER) is defined to unify the performance of heterogeneous SC3 functions (sensing, comms, compute, control) into a single information-theoretic measure.
LOAC Framework: Development of a hybrid Learning-Optimization framework that achieves near-optimal solutions with low computational complexity. It effectively bridges the gap between the speed of learning-based methods and the optimality of mathematical optimization.
Theoretical Proof: Proof that the resource allocation sub-problem is convex when pairing is fixed, guaranteeing the Critic's evaluations are globally optimal for any given pairing.

4. Simulation Results

The authors evaluated the LOAC framework against several benchmarks in a disaster rescue scenario (20 sensors, 4 actuators, 6G airborne EIH):

vs. Exhaustive Search: LOAC achieves near-optimal LQR costs (matching the performance of exhaustive search) but with polynomial-time complexity, whereas exhaustive search suffers from factorial complexity, making it infeasible for real-time use.
vs. Optimization-Based Schemes (Relaxation): Traditional methods that relax integer variables to continuous ones and use rounding suffer significant performance loss due to the loss of intrinsic coupling. LOAC outperforms these by preserving the discrete nature of pairing.
vs. Heuristic Pairing (Comm-first, Compute-first, CCA):
- Comm-first works well in bandwidth-limited regimes but fails when computing becomes the bottleneck.
- Compute-first works well when bandwidth is abundant.
- LOAC dynamically adapts, automatically balancing communication and computing capabilities to prevent bottlenecks, consistently outperforming static heuristic strategies.
vs. Communication-Oriented Schemes: Schemes maximizing sum-rate or min-rate (without direct control cost consideration) perform poorly. LOAC, by directly optimizing for LQR cost, significantly reduces control error.
Convergence: The DNN converges rapidly (within ~1500 epochs) even under extreme bandwidth conditions, demonstrating the efficiency of the Critic's feedback mechanism.

5. Significance and Impact

Paradigm Shift: The paper shifts the focus from "network-centric" optimization (maximizing throughput) to "task-centric" optimization (minimizing control cost), which is crucial for 6G autonomous systems.
Real-Time Viability: By decoupling the combinatorial pairing from continuous resource allocation via the LOAC framework, the solution becomes computationally feasible for dynamic, time-varying environments like disaster zones.
Holistic Design: It demonstrates that treating the SC3 loop as an integrated entity (rather than separate links) is essential for maximizing system utility. The dynamic bidirectional adaptation between task requirements and resource availability is key to robust autonomous operations.

In conclusion, this work provides a robust, scalable, and theoretically grounded framework for orchestrating 6G-enabled autonomous systems, ensuring that limited communication and computing resources are allocated precisely where they are needed to maintain stable and efficient control loops.