【Control】🧠 03. PID Control Is More Complete Than You Think

topics: [“control engineering”, “PID”, “AI”, “LLM”, “robotics”]

Introduction

As AI and LLM-based control attract increasing attention,
many people—including myself—tend to think:

• “Traditional control theory may be approaching its limits.”
• “AI could enable much smarter control.”

With that expectation, I designed and implemented an
AITL (Adaptive / Intelligent Triple-Layer) control architecture,
combining PID × FSM × LLM, and evaluated it against a standalone PID controller
using Python-based simulations.

The conclusion, stated plainly, was this:

Modern control theory—especially PID control—is far more complete and robust
than commonly assumed, and even with AITL, it is difficult to demonstrate
a clear advantage over a well-designed PID controller.

This article documents that evaluation process
and the sober conclusions drawn from it.

🔍 Why I Tried to Build AITL

The idea behind AITL is straightforward:

• PID: real-time control (inner loop)
• FSM: state supervision and mode switching (middle loop)
• LLM: redesign, re-identification, and decision support (outer loop)

The underlying expectations were:

• “PID is fast but rigid.”
• “FSM detects abnormal conditions.”
• “LLM redesigns the controller instead of a human.”

In particular, I expected benefits in systems with:

• Large disturbances
• Long-term operation with changing characteristics

🎯 Baseline: Implementing PID Properly

A critical point in this evaluation was not to trivialize PID control.

The baseline PID controller was implemented with:

• Full P / I / D terms enabled
• Explicit disturbance inputs
• Evaluation including divergence and saturation conditions
• No artificial handicaps (e.g., setting I = 0)

Under these conditions, a properly tuned PID controller showed:

• Robust disturbance rejection
• Fast response
• Clearly explainable behavior
• Easy debugging and maintenance

In short, a well-designed PID controller is extremely strong.

🧩 Adding AITL (PID × FSM × LLM)

Next, FSM and LLM layers were added:

• FSM: detects increased disturbances or degraded response
• LLM: generates redesign strategies for PID gains
• Re-identified PID is re-applied

Structurally, this is sound.
It simply separates what humans already do.

However, in practice, several issues emerged:

• Redesign introduces time delays
• Design intent becomes opaque (black-box behavior)
• Performance differences compared to PID alone
  are not clearly visible in response plots

📉 Results: No Clear Advantage

To be completely honest:

Under these conditions, no clear win over standalone PID was observed.

From both performance and operational perspectives:

• Stability: PID alone was sufficient
• Responsiveness: PID was faster
• Explainability: PID was overwhelmingly superior
• Maintainability: PID was far more practical

🧠 Why PID Is So Strong

The reasons are straightforward:

• The theory is mature
• Decades of field experience are embedded
• Robustness is a first-class design goal
• Behavior is mathematically explainable

A properly designed PID controller is already close to optimal
—that was the strongest takeaway from this study.

🌱 Where AITL Might Still Make Sense

This does not mean AITL is meaningless.

AITL may be valuable in contexts such as:

• Unmanned or remote systems
• Long-term autonomous operation (space, deep sea, etc.)
• Environments where frequent human redesign is impossible
• Layers outside direct control performance (decision and reconfiguration)

In other words:

AITL is not a technology to directly improve control performance,
but a framework for structuring design and decision-making.

🧾 Conclusion

AI- and LLM-based control is appealing.
But once implemented and compared honestly:

Modern control theory is already very strong.

This is not a failure.
It is the result of testing reality against expectation.

AITL should not be positioned as a replacement for PID,
but rather as a conceptual scaffold to revisit
when truly necessary.

📝 Closing Remarks

“AI will dramatically improve everything”
—reality is not that simple.

But:

Trying it, and drawing a clear boundary,
is itself a meaningful engineering outcome.

I hope this serves as a useful reference
for others facing similar questions.