I am intrigued to comprehend the difference between Adaptive Controls and Machine Learning? Are these derived from the same concepts? Is feedforward/feedback control the same as reinforcement learning?
Adaptive controllers mean the parameters of the controller are updated/modified for every set-point/operating condition. This adaptation will be done using fuzzy logic/neural networks/optimization algorithms etc.
Refer to https://ieeexplore.ieee.org/abstract/document/9304405 and Prof. Anuradha Annaswamy's publications (from MIT)
Dear Mihir Malladi,
- For adaptive control, the optimality or optimization objectives are defined nominally beforehand, while for machine learning techniques, the optimization objectives are a function of the learning process (evolutionary) and are expressed in general terms. A tracking error must converge quickly to its minimum value, whereas it is allowed to fail in the learning phase.
- There is a difference between the way of proceeding and the way of arriving at the expected goal, for the adaptive control the computation time is a fixed horizon while for the ML the horizon is running until a convincing result is reached.
- Their intrinsic conception is different, one is purely analytical and the other is meta heuristic.
However, we can combine the two techniques to benefit from the best of each.
For more details and information about this subject, I suggest you see the links on the topic.
https://cs.stackexchange.com/questions/32163/whats-the-difference-between-adaptive-control-and-hierarchical-reinforcement-le
https://acc2021.a2c2.org/presentation/plenary/lessons-adaptive-control-towards-real-time-machine-learning
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.452.9944&rep=rep1&type=pdf
https://deepai.org/publication/connections-between-adaptive-control-and-optimization-in-machine-learning
Best regards
Your question is non-trivial. It is reflected in answers you got: they are
only partial and as such of a limited validity. Importantly, they neglect long history and achievements of both domains. I will try to answer a bit better but a deeper understanding would require to inspect and compare
results summarised in works of big authors as Astrom, Bellman, Bertsekas, Haykin, Hutter, Feldbaum, Simon, Peterka, etc.
The discussed domains overlap. Both incorporate a part that learns model of environment (controlled system). Adaptive control explicitly aims to influence properties of the closed loop formed by the environment and controller (agent, decision maker). A deeper inspection of theoretical background of both domains shows that ML can be strictly embedded into the general adaptive control theory. The source communities (computer scientist vs. control or decision-making researchers) bring main differences in language, aims and stress. For instance, control cares about closed-loop stability, stresses on-line implementation (unlike ML) and cares much less about computational complexity and algorithmic ways how to fight with curse of dimensionality ....
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Writing