Are Control Theories (Control Laws) Becoming Obsolete?
A New Paradigm
Is basic control theory obsolete? the answer is Yes and No: Yes, because AI driven models have proven to easily surpass most human based expert control method, in performance and in conceptualization time. in other words, and AI agent can design a better control method in a fraction of a the time it takes the control system expert.
Alternatively: No, because Experience, is still needed to understand the physics of it: you must have a background in classical control to understand what the physics of the system can do. Such know how is needed to define the Cost Function Engineering and defining the Reward functions. Designing the data sets, qualifying the data quality,
understanding the frequency content, sorting out all the variables, and their respective relevancy and reliability. Finally, observing how the control is failing and fill in new data to help the AI agent requires expert data corrections.
Importance of the Data
One of the key disadvantages of AI based control method is the criticality of the data, the saying goes:
Bad Data + <Anything> = Bad Performance
it is not the quantity of the data that matters, but rather the quality of the data. In this world we have two main sources of data: real data from the actual physical system and simulated data from a computer model (digital twin). The former is generally qualified as high quality data but is often limited and expensive in nature, the latter is of lower quality data but cheap and plentiful.
Practical Implications
One concept that one to understand is that you cannot improve/modify your physical system without invalidating your data. This has two implications:
First, an AI policy for a control system cannot be tuned; it has already been tailored by the AI agent to provide the best compromise based on the target goal. the only way to improve an AI based control system performance is to identify the domains of underperformance and add training data to improve the control policies at that specific underperformance area.
Second, a hardware modification to the system will require a new control policy, and a new set of training data, therefore the engineering pipeline for collecting data and training the control policies has to be efficient and optimized such that a new control policy can be issued in a timely fashion, despite the large amount of data typically involved in the training.
Research Directions
As defined above the importance of having a "fast science" data production and data engineering pipeline is critical. It means that data collection training and evaluation must be one continuous and efficient operation. At the present time, edge computing is only capable of executing AI inference, data generation (such as from a digital twin) and training are task that belong to powerful GPU based servers. it means that the control policy production is distributed and multilayer engineering effort. In that context we can see envision that the edge real time controller should be designed to facilitate the "fast science" philosophy, it should be designed for Agentic AI, Cooperative control and In-the-loop simulation. this new concept of in the loop simulation allows for a hybrid data production method that maximizes quantity and quality of the training data.