Learning algorithms can enhance the adaptive ability of the control system, but they tend to require iterative calculation, which may influence the real time performance .
Think about the idea in the conceptual and practical levels.
Lets think about the idea in the conceptual level: what is the difference between this idea and adaptive control strategies. Adaptive controllers conceptually do the same; they have (1) a basic real time controller which guarantees the stability of the system and (2) a learning part (adaptive term) which improves the performance of the overall controller by iteration.
In the application level, I think it makes sense. Actually, the learning algorithms require more trials (than adaptive ones) to converges to the optimum solutions. And maybe a novel "combination" of a basic (non adaptive controller) controller with a learning algorithm (as imitation learning or Q-leaning) leads to an effective behavior and convergence performance.
Considering adaptive control the most basic form of "learning": yes, learning is feasible in real-time control systems.
When it comes to more mature learning, you might have to pay attention to your algorithms: your control system always has to provide the runtime to execute the algorithm. But it is possible to implement the learning pat in a less critical thread than e.g. the data acquisition providing the input for the learning.
Short answer: yes, it is practical. It is actually a science called 'reinforcement learning (RL)' and is inspired by the way the brain learns by practice. There are different approaches to do RL ranging from good old Markov Decision Process and Q-learning to more renovated ones like Deep Q-learning Neural networks (DQN).
- Badges
- Science topic