📊 statistics

Interpretable PID Parameter Tuning for Control Engineering using General Dynamic Neural Networks: An Extensive Comparison

This paper demonstrates that General Dynamic Neural Networks (GDNN) can effectively extend PID controllers to handle complex, non-linear systems with superior performance and noise robustness compared to standard and model-based approaches, while simultaneously ensuring interpretability and stability through bounded-input bounded-output analysis.

Original authors: Johannes Günther, Elias Reichensdörfer, Patrick M. Pilarski, Klaus Diepold

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Johannes Günther, Elias Reichensdörfer, Patrick M. Pilarski, Klaus Diepold

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 you are driving a car. In the old days, the cruise control was a very simple, rigid system. If you set it to 60 mph, it would try to keep the speed at 60 mph no matter what. If you went up a steep hill, the car would slow down because the simple system didn't know how to push harder. If you went down a hill, it might speed up too much. This is like a standard PID controller (Proportional-Integral-Derivative), which is the "workhorse" of modern automation. It's great for simple, predictable tasks, but it struggles when the world gets messy, bumpy, or unpredictable.

This paper proposes a solution: giving that simple cruise control a "brain" in the form of a Neural Network.

The Problem: The Rigid vs. The Complex

The authors point out that while our factories and machines are becoming incredibly complex (like driving a car through a storm while the road is changing), we are still mostly using those simple, rigid controllers.

The Old Way: Engineers tune the controller once before the machine starts. It's like setting your cruise control and hoping the road stays flat.
The Issue: If the road changes (a disturbance) or the car gets noisy (sensor noise), the simple controller fails or needs a human to jump in and fix it.
The New Idea: What if the controller could "think" and adjust its own settings while it's driving?

The Solution: The "Smart" Cruise Control

The researchers built a system where a small Neural Network (a type of AI) sits inside the controller. Instead of just following fixed rules, this AI watches the car's speed and the error (how far off the target speed it is) and instantly tweaks the controller's three main knobs (the PID parameters) to handle whatever is happening right now.

Think of it like a chameleon. A standard controller is a lizard that stays the same color no matter what. This new "Neural PID" is a chameleon that instantly changes its color (adjusts its settings) to match the background (the current situation), whether it's a sunny day, a rainy night, or a bumpy road.

The Test Drive: Four Tough Challenges

To prove this works, the authors didn't just test it on one easy task. They put their "Smart Cruise Control" through four very different, difficult driving scenarios (benchmarks):

The Two-Tank System: Imagine trying to keep the water level in a lower tank perfect, even though water flows from a top tank in a messy, non-linear way. It's like trying to fill a bucket with a hose that has a weird kink in it.
The Inverted Pendulum: This is the classic "balancing a broomstick on your hand" problem. It's naturally unstable; if you let go, it falls. The controller has to push the cart back and forth perfectly to keep the stick upright.
The Delayed System: Imagine driving a car where there is a 5-second delay between you turning the steering wheel and the car actually turning. This is a nightmare for standard controllers.
The Chaotic Loop: This is like trying to control a fluid that is swirling in a completely unpredictable, chaotic pattern, like a whirlpool.

The Results: The AI Wins Most Races

The authors ran these tests 50 times each, adding "noise" (static on the radio) and "disturbances" (sudden gusts of wind or potholes).

The Scorecard: In 15 out of 16 different scenarios, the Neural PID controller did a better job than the standard, rigid PID controller.
Beating the Experts: It even beat specialized, complex mathematical controllers (like "Backstepping" or "Smith Predictors") in 13 out of 16 scenarios.
The "Unstoppable" Moment: In the chaotic loop test, when a sudden disturbance hit, the standard controllers went crazy, oscillating wildly and using too much energy. The Neural PID, however, stayed calm, adjusted its settings on the fly, and stabilized the system.

The "Black Box" Problem: Making it Safe

Here is the biggest hurdle. Engineers are scared of AI because it's a "black box." They don't know why the AI made a decision. If an AI controller crashes a plane, no one knows if it was a bug or a feature. In safety-critical industries, you can't just trust a black box.

The authors solved this by adding a Stability Check.

The Analogy: Imagine the AI is a pilot. Usually, we don't know if the pilot is flying safely until we look at the destination. The authors built a "co-pilot" that constantly checks the math to ensure the plane will land safely, no matter what the AI pilot is doing.
The Result: They showed that even though the AI is changing the settings every millisecond, they can prove mathematically that the system remains stable. This turns the "black box" into a "glass box" that engineers can understand and trust.

The Bottom Line

This paper says: "We can make simple controllers super-smart by giving them a small AI brain that adjusts itself in real-time. We tested it on the hardest problems in engineering, and it won almost every time. Plus, we figured out a way to prove it's safe, so engineers can actually use it in the real world."

It's a step toward machines that don't just follow orders, but adapt to the chaos of the real world, all while keeping a safety net that humans can understand.

Technical Summary: Interpretable PID Parameter Tuning for Control Engineering using General Dynamic Neural Networks

Problem Statement
Modern automation systems increasingly involve complex physical processes that cannot be adequately modeled by simple linear equations. While Proportional-Integral-Derivative (PID) controllers remain the industry standard due to their simplicity and ease of tuning, they are inherently limited to linear or near-linear systems. When applied to complex, non-linear processes, standard PID controllers often fail to provide robustness against disturbances or varying system dynamics, necessitating close human monitoring. Conversely, advanced control methods designed for non-linear systems are often too complex for widespread engineering adoption.

While machine learning techniques, specifically neural networks, have shown potential in tuning PID parameters online to adapt to non-linearities, their application in real-world control engineering is hindered by two primary barriers:

Lack of Generalization: Previous studies have typically focused on specific, single-task architectures (e.g., strictly recurrent or diagonal networks) that may not generalize across diverse control environments.
Lack of Interpretability and Safety: Neural networks function as "black boxes," preventing engineers from verifying the stability of the suggested parameters. Control engineering demands rigorous stability guarantees (specifically bounded-input bounded-output stability) to ensure safety, a requirement often ignored in existing neural tuning literature.

Methodology
The authors propose a framework that integrates General Dynamic Neural Networks (GDNNs) into a standard closed-loop control architecture to perform online PID parameter tuning.

Architecture: The neural network receives the system output ( $v$ ) and the control error ( $e$ ) as inputs. It outputs the three tunable PID constants ( $K_P, K_I, K_D$ ) at each time step. To ensure computational efficiency for hardware with limited resources (demonstrated on a Raspberry Pi 3), the networks are restricted to a single hidden layer with four neurons. The architecture includes recurrent connections where neuron outputs are fed back with a one-time-step delay ( $g^{-1}$ ).
Training Strategy: Unlike standard deep learning approaches that rely on end-to-end optimization with known ideal outputs, this framework cannot use standard backpropagation because the "ideal" PID parameters are unknown. Instead, the authors employ a numerical Levenberg-Marquardt algorithm. This method approximates the Jacobian matrix of the neural network weights using finite differences to minimize the squared control error.
Experimental Design: The study evaluates the approach across four distinct benchmark systems representing common control engineering challenges:
1. Two-Tank System: A non-linear system representing processes like bio-reactors.
2. Inverted Pendulum on a Cart: A system with unstable equilibrium and non-linearity.
3. System with Non-negligible Time Delay: A first-order LTI system with significant delay.
4. Chaotic Thermal Convection Loop: A system exhibiting chaotic behavior.
Robustness Testing: Each system is tested under four conditions: without noise/disturbance, with noise only, with disturbance only, and with both. Disturbances are applied to push the system to the limit of stability. The neural PID is compared against a standard PID controller and a system-appropriate model-based controller (e.g., Backstepping, LQR, Smith Predictor).

Key Contributions

Extensive Comparative Study: The paper advances the state of the art by moving beyond single-task evaluations. It provides a comprehensive comparison of GDNN-based PID tuning across four diverse benchmark systems, including scenarios with sensor noise and external disturbances.
Interpretability via Stability Analysis: The authors address the "black box" concern by applying bounded-input bounded-output (BIBO) stability analysis to the neural network's output. By computing the closed-loop transfer function at every time step based on the neural network's suggested parameters, they demonstrate how to verify system stability. This is illustrated using the chaotic thermal convection loop, showing that the system becomes input-output stable as it approaches the steady state, making the neural network's decisions understandable to control engineers.

Results
The study, based on 50 independent runs per scenario (16 total scenarios), yielded the following quantitative results:

Performance: The neural PID controller outperformed the standard PID controller in 15 out of 16 tasks.
Comparison to Model-Based Control: The neural PID controller outperformed the specialized model-based controllers (Backstepping, LQR, Smith Predictor) in 13 out of 16 tasks.
Specific Findings:
- In the Inverted Pendulum scenario, the neural PID was the only controller capable of stabilizing the system when subjected to a significant disturbance, whereas standard PID and LQR failed (resulting in error values of 140 vs. 0.27).
- In the Chaotic Thermal Convection Loop, the neural PID achieved a superior balance between settling time and overshoot compared to Backstepping and standard PD. It also demonstrated significantly lower energy consumption during stabilization.
- In the Time Delay system, the neural PID outperformed the standard PID in all scenarios, though the Smith Predictor (which has explicit knowledge of the delay) performed slightly better in noise-free conditions.

Significance and Claims
The paper claims that this work represents a significant step toward the acceptance of machine learning in real-world control systems. By combining the performance benefits of neural networks with rigorous stability analysis, the authors demonstrate that neural PID controllers can be both high-performing and interpretable.

The authors modestly frame their work as a "first step" and a "first investigation" using general dynamic neural networks for this purpose. They emphasize that while the training data was derived from differential equations, the resulting controllers showed increased resilience to unforeseen disturbances and noise. The primary significance lies in bridging the gap between the adaptability of machine learning and the safety requirements of control engineering, thereby addressing a key barrier to the deployment of AI in safety-critical systems. The paper does not claim to have solved all safety issues but suggests that incorporating stability constraints (e.g., as regularization terms) into the training process is a natural and necessary extension for future work.