Machine learning or PID?

Question

Machine learning

Both algorithms are closed-loop and rely on target values, but which is better? Let's say we have a platform (or an airplane, it's easier) that needs to be stabilized in space. For this, we have a gyroscope and three axes - tonnage, roll, yaw. The sensor readings go to the input of the neural network or the PID controller, and at the output they generate control signals for the servomotors, the target is set to zero corresponding to the equilibrium of the platform and various forces act on it (in the case of an airplane, wind, turbulence, etc.).

Answer 1

Some perfect porridge. The fact that a shovel and a rake have almost the same design - a stick with a piece of iron at the end - does not mean that they can be compared. The fact that both approaches use the feedback principle does not mean that they are at least somewhat similar and we can talk about which is better and which is worse.
Start over.
PID is a classic example of an application from control theory that is about a hundred years old. In fact, when you drain the water in the toilet tank, a feedback loop starts working for you: until the water is filled, the valve is open, the water flows. The water raised the float (sensor) - the lever tilted (transmission link) - the inlet valve closed (actuator) - the water stopped flowing. In your platform, in fact, the same thing, only instead of a float - gyro sensors, instead of different water pressure - "wind, turbulence, etc.", instead of a valve - a servomotor. And where is the place for machine learning and neural networks?
Machine learning and neural networks are not an object management tool. These are tools for building an object model. In a huge field of knowledge, such models are built on the basis of physical and mathematical models (mechanics, electricity, thermal conductivity, aerodynamics, radio communications, etc., etc.) and applying machine learning there is like digging with a rake. There, the models are clear how to build and how to use - well, why a neural network to build a model of an electric iron, if Ohm, Kirghof and Joule have already chewed everything up a long time ago.
But there are many areas where there are no such laws. They either do not exist in principle (medicine, marketing, psychology), or they do exist, but the real conditions are so “noisy” that laws are useless in a refined “school” form, and it is simply impossible to take into account all interfering factors (well, for example, wear and tear mechanism or the occurrence of malfunctions in complex technical objects). And this is where machine learning begins, the essence of which is in the slogan "give us the data, we will derive the formulas ourselves, we often cannot and will not explain them, but the resulting models will very accurately describe real processes." Well, for example - face recognition, games, economic consequences, etc.
But, most importantly, machine learning is not a management process, it is a process of building a model! When they talk about neural networks, they mean, first of all, their training, which really (coincidence) uses the feedback principle, i.e. builds a model, adjusting it to the process. But then, to solve a specific problem, the RESULTS of training the neural network are taken, that is, in fact, some logical and mathematical model and it is applied to solve the problem of either control, or prediction, or diagnostics.
Something like that.