💻 computer science

FuMO-Edge: Interval Type-2 Fuzzy Logic with NSGA-II for Multi-Objective Workload Orchestration in Mobile Edge Computing

This paper proposes FuMO-Edge, an online mobile edge computing orchestrator that integrates Interval Type-2 fuzzy logic with NSGA-II to optimize task placement under uncertainty, demonstrating significant reductions in service time and failure rates compared to existing type-1 baselines, particularly in high-density device scenarios.

Original authors: chunmao jiang, bo sun, ruyi ye, hao zhang

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: chunmao jiang, bo sun, ruyi ye, hao 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

Imagine a city where thousands of people are trying to use their phones at the same time, each asking for a different service: some need instant video calls, others are running complex health monitors, and many are downloading large files. In the old model of computing, all these requests would travel far away to a massive, centralized data center to be processed. But as our devices have become faster and our needs more urgent, that long journey has become too slow. The solution is to move the computing power closer, placing servers right at the edge of the network, perhaps in a local cell tower or a neighborhood hub. This is mobile edge computing. It brings the brain of the internet closer to the hand, but it creates a new, difficult problem: how does the system decide, in a split second, whether to handle a task right there at the edge, send it to a nearby server, or push it all the way to the distant cloud?

The decision is complicated because the environment is messy and unpredictable. Wireless signals fade, servers get overloaded, and people move around, changing the network landscape in real time. If the system guesses wrong, a video call might freeze, a medical alert could be delayed, or a task might fail entirely. For years, researchers have tried to solve this using two main approaches. One uses fuzzy logic, a way of making decisions based on vague, human-like rules such as "if the network is slow, send it to the cloud." The other uses evolutionary algorithms, which are like digital natural selection, where thousands of potential solutions compete and evolve to find the best one. However, the fuzzy logic methods have been limited because they rely on fixed, hand-written rules that cannot account for the uncertainty of the moment, while the evolutionary methods often assume the world is perfectly predictable, which it never is.

A team of researchers at Fujian University of Technology has bridged this gap with a new system they call FuMO-Edge. Instead of choosing between a rigid rulebook and a blind search, they combined the two. They took the fuzzy logic approach and upgraded it to handle uncertainty directly, then used the evolutionary algorithm to teach the system how to set its own rules. The key innovation is a concept called the "footprint of uncertainty." In a standard fuzzy system, a rule might say "if the load is high," but it doesn't know how much "high" can vary. The new system adds a fuzzy band around that definition, allowing the decision-maker to say, "I know the load is high, but I am also aware that my measurement might be slightly off, so I will act cautiously." This band is not fixed; the researchers let the evolutionary algorithm adjust the size and shape of this uncertainty band for every single rule, essentially teaching the system how much to trust its own sensors in different situations.

To test this, the researchers built a massive simulation involving up to 2,400 mobile devices generating four different types of workloads, from light health monitoring to heavy data processing. They ran the simulation thousands of times, comparing their new system against seven other methods, including older fuzzy systems and standard evolutionary approaches. The results showed that in crowded, high-stress environments, the new system significantly outperformed the others. At the highest density of devices, it reduced the time it took to complete tasks by 31 percent and cut the rate of failed tasks by 29 percent compared to the best existing fuzzy method. Perhaps more importantly, the system was much more stable; when the researchers ran the test multiple times with different starting conditions, the new method consistently delivered good results, whereas the older methods sometimes failed spectacularly.

The study revealed exactly where this improvement came from. By analyzing the rules the system learned, the researchers found that the "footprint of uncertainty" grew larger for the most chaotic inputs, such as the overall load on the network or the size of the data being sent. The system learned to be more flexible and cautious when the network was congested, preventing it from making rash decisions that would lead to failure. Conversely, for inputs that were easier to measure directly, like the speed of a specific connection, the system kept its rules sharp and precise. This ability to adapt its own level of caution based on the situation allowed it to navigate the chaos of a busy network far better than systems that relied on fixed rules or assumed perfect information.

It is important to note that these findings come from a sophisticated computer simulation, not a physical test with real phones and towers. The researchers used a well-established simulation platform that mimics the behavior of mobile networks, but they did not measure the system on actual hardware. While the results are promising and statistically significant within the simulation, the authors acknowledge that real-world conditions, such as sudden bursts of traffic or hardware failures, could behave differently. They also point out that their system currently uses a simplified measure of energy usage rather than a detailed power model, and that the time it takes for the system to make a decision in real-time has not yet been measured on physical devices.

Despite these limitations, the work demonstrates a clear path forward for managing complex networks. The researchers showed that by teaching a system to understand its own uncertainty, rather than just following a static set of instructions, it becomes much more resilient. The system does not need to know the future to make a good decision; it simply needs to know how unsure it is about the present. As mobile networks continue to grow and become more critical to daily life, this ability to adapt to the unknown may become the difference between a smooth experience and a broken connection. The code and data from this study have been made available to other researchers, inviting further exploration into how artificial intelligence can learn to manage the chaotic, unpredictable flow of our digital world.

Technical Summary: FuMO-Edge

Problem Statement
Mobile Edge Computing (MEC) orchestrators face a critical challenge: deciding where to execute tasks (nearest edge, remote edge, or cloud) under significant uncertainty. This uncertainty stems from fluctuating wireless channel quality, heterogeneous and time-varying edge server loads, and user mobility. Existing solutions suffer from specific limitations:

Fuzzy-logic orchestrators (e.g., Sonmez et al., 2019) typically use Type-1 inference with hand-tuned membership functions. While explainable, they cannot quantify "second-order uncertainty" (uncertainty about the membership functions themselves) and rely on manual rule construction.
Multi-objective evolutionary algorithms (MOEAs) like NSGA-II generate Pareto-optimal task assignments but generally assume deterministic inputs, treating uncertainty only via post-hoc averaging rather than embedding it into the decision logic.
Hybrid approaches exist but have not yet combined Interval Type-2 (IT2) fuzzy inference with NSGA-II specifically for online MEC task offloading, nor have they exploited the Footprint of Uncertainty (FOU) to absorb runtime input noise.

Methodology: FuMO-Edge
The authors propose FuMO-Edge, an online MEC orchestrator that integrates Interval Type-2 (IT2) fuzzy logic with NSGA-II multi-objective optimization. The approach follows a two-stage process:

Offline Training (NSGA-II):
- Gene Encoding: The system evolves an 84-dimensional vector representing the parameters of 21 triangular IT2 membership functions (7 input variables × 3 linguistic terms). Each function is defined by four genes: the vertex triple $(a, b, c)$ and the FOU half-width $\delta$ .
- Validity by Construction: The encoding ensures that any point in the search space decodes to a valid IT2 set without requiring repair operators. The FOU is symmetric around a shared apex.
- Fitness Evaluation: NSGA-II uses a "simulator-in-the-loop" approach. Each candidate gene vector is evaluated by running a full EdgeCloudSim simulation (1200 devices, 10 minutes).
- Objectives: The optimization minimizes three conflicting objectives:
  1. Average Service Time ( $F_1$ ).
  2. Task Failure Rate ( $F_2$ ).
  3. A weighted Edge-Cloud utilization proxy ( $F_3$ ), serving as a resource-cost metric.
- Selection: From the resulting Pareto front, a single "knee-point" solution is selected using a TOPSIS-style rule (closest to the ideal point) to serve as the deployed policy.
Online Deployment:
- The selected gene vector instantiates an IT2 fuzzy inference engine.
- At runtime, the orchestrator uses the Karnik–Mendel (KM) type-reduction algorithm to handle the IT2 sets, making real-time offloading decisions for every incoming task based on current network and load measurements.
- The rule base structure (81 + 27 rules) and rule consequents are inherited from the Sonmez et al. (2019) baseline and remain fixed; only the input membership functions are evolved.

Key Contributions
The paper outlines four primary contributions:

C1. Evolvable IT2 Membership Functions: Unlike prior IT2 edge work that derives FOU from peer statistics, FuMO-Edge evolves the FOU half-widths ( $\delta$ ) as part of the Pareto optimization, allowing the system to learn the optimal level of uncertainty absorption for specific inputs.
C2. Simulator-in-the-Loop Tuning: The membership functions are optimized directly against deployment-relevant outcomes (service time, failure rate) via EdgeCloudSim, rather than a surrogate loss function.
C3. Controlled Comparative Study: The authors benchmark FuMO-Edge against seven baselines, including classical heuristics, hand-tuned Type-1 fuzzy orchestrators, a Type-1 NSGA-II ablation (where $\delta=0$ ), and a fuzzy-free NSGA-II. The study spans 12 device densities (200–2400), 30 iterations, and four application workloads, totaling 2,880 simulation runs.
C4. High-Density Empirical Gains: The study demonstrates that the benefits of the FOU are concentrated in high-density regimes where input uncertainty is highest.

Results
Evaluated at the highest density (2,400 devices), FuMO-Edge achieved the following improvements compared to the primary Type-1 baseline (Sonmez 2019):

Service Time: Reduced by 31% (4.15s $\to$ 2.86s).
Failure Rate: Reduced by 29% (13.96% $\to$ 9.92%).
Utilization Proxy: Reduced by 9%.

Ablation and Mechanism Analysis:

Role of FOU: When the FOU was locked to zero (T1-NSGA-II ablation), the failure rate increased to 16.62%, worse than the hand-tuned Type-1 baseline. This confirms the FOU is the principal mechanism for failure-rate reduction.
Input Sensitivity: The evolved FOU was not uniform; it concentrated uncertainty on congestion-sensitive inputs (average edge utilization, task size, delay sensitivity, remote edge utilization) while keeping directly sensed inputs (nearest edge utilization) relatively crisp. Notably, the blur for 'wan bw' varied widely across independent training seeds and was not identified as a stable mechanism.
Workload Specificity: The gains were most pronounced for heavy, bandwidth-bound workloads (Heavy Computation, Infotainment). Light workloads (Health Monitoring) showed negligible difference between IT2 and Type-1.
Stability: Across five independent NSGA-II seeds, FuMO-Edge maintained a lower and tighter failure rate distribution (9.4% $\pm$ 1.6%) compared to the Type-1 ablation (15.6% $\pm$ 3.5%), whereas the fuzzy-free NSGA-II was highly unstable (21–87% failure).

Significance and Claims
The paper claims that FuMO-Edge successfully bridges the gap between fuzzy logic and evolutionary optimization for online MEC orchestration. Its primary significance lies in demonstrating that evolving the Footprint of Uncertainty allows the orchestrator to absorb runtime input noise effectively, leading to significant reductions in task failure rates specifically under high-load, high-uncertainty conditions.

The authors explicitly note that:

The results are simulation-based (EdgeCloudSim) using synthetic Poisson workloads; physical validation on real testbeds and real traffic traces is required for future work.
The "energy" objective is a utilization proxy, not a calibrated power model.
The online decision latency overhead of IT2 type-reduction is analyzed theoretically ( $O(R)$ ) but not empirically measured on target hardware.
The rule consequents were kept fixed; co-evolving them is a suggested future direction.

The study concludes that while Type-1 fuzzy logic and NSGA-II are effective individually, their combination with an evolved Interval Type-2 representation provides a robust mechanism for handling the specific uncertainties inherent in dense mobile edge environments.

Technical Summary: FuMO-Edge

More like this