From PID to Neural Network Driven AI Control Policies
PID control, the legacy all purpose tool
PID controllers are mostly linear, manually tuned algorithms. they are effective in stable environments but struggle with the complexity and variability of most modern control situations. PID cannot optimize toward multiple goals unless you accomodate custom tunning for different setpoint. PID is best suited for single input- Single output (SISO) systems and for anything but the simplest situation multiple PIDs loops are often required.
Limiting factor: tuning the system faces multiple trade-off
- Time to setpoint
- Stability (oscillations)
- Non-linear artifacts (valve, conductance)
- System dynamic variability (gas flow)
Model Predictive Control (MPC),
MPC is beneficial in feedback and feedforward but still lack the capability to model the complexity and variability of some complex systems. it represent a valuable improvement upon PID particularly to help in real time parametrization but requires a reliable process models. aditionally solving the model, in real-time, is a mathematical optimization problem that requires a lot of compute.
Limiting factor: difficult to create a reliable model at the edge
- The high speed dynamics (unknown)
- Difficulty with model reduction (order and sampling)
- General robustness of the model is weak unless it is very simple
Agentic AI - Machine Learning (ML)
Deep learning models, handle highly complex, high-dimensional data and handle relationships between variables that are intricate and nonlinear. Neural Networks (NN) are used as mapping functions from high dimensional data to control actions. NN can optimize control performance toward multiple goals.
Limiting factor: not an evolution, completely different approach to real time control
- Requires large volume of real time operational data
- The control policy is tunned for a specific data context
- Understanding of goal-based logic require new thinking
- Engineering pipeline setup is initially complex