I have recently started doing research on Machine Learning (ML) and Reinforcement Learning (RL) and their applications in the field of mechanical engineering. I am also interested to know more about the available resources, more specifically any books, to start with.
Thank you
Article Learning event-triggered control from data through joint optimization
and more, for robotics and intelligent transportation; interested to merge efforts/jointly to generate value within the discourses/branches...
Chemistry:
RL can also be applied in optimizing chemical reactions.
Robotics:
There are tremendous work on applying RL in Robotics.
Traffic Light Control:
researchers tried to design a traffic light controller to solve the congestion problem.
Thank you, Hossam Elzayady
Do you have any suggestions for someone who wants to start learning RL?
Hi Nima Geran Malek
This is my opinion in solid mechanics:
- Machine learning and Neural Network can be used to reduce the computational cost of a computational expensive simulation:
Imagine that we verified and experimentally validated a numerical simulation. But the simulation is very computationally expensive. To optimize the structure we may need to run this simulation iteratively. So in this case, regression (in machine learning methods or NN) can be employed to achieve a continuous regression function and then we optimize that function with the gradient-based method.
(For reference, there is a good book to explain ML in civil engineering that its concepts is very close to mechanical engineering in a sense of modeling; Goulet, James-A. Probabilistic machine learning for civil engineers. MIT Press, 2020.)
- Reinforcement learning I think paves the way to do Structural Health Monitoring in which we need to detect online any failure in structure in real-time. There is good literature about this topic.
Nevertheless, although Machine Learning, Neural Network, and Reinforcement learning are very hot topics these days, there is a lack of physical sense in these methods and their prediction capability only limited to the input data (and some hyperparameters). Therefore, if we want to use these methods, I think we should carefully verified and validate the input data. More importantly, we should also compare their predictions with experimental measurements.
Dear Nima Geran Malek i'm conquering with Hossam Elzayady answer and you can start using anaconda as an environment for RL or use online python.
Google Colab is a free cloud service and now it supports free GPU and TPU . you can use Colab an environment for RL. Most libraries you need are installed and implemented on google colab. Ashraf Abosekeen Nima Geran Malek
https://adventuresinmachinelearning.com/reinforcement-learning-tutorial-python-keras/
Dear Nima Geran Malek as the latest applications of Reinforcement Learning (RL) in the field of engineering I suggest to your attention the WuDao 2.0 model of China Neural Network (NN) with 1,75 trillion parameters. Such neural network for the first time demonstrated multitasking. It can not only process and generate texts but also analyze and synthesize images, including the input texts and vice versa.
https://www.researchgate.net/post/Is_it_China_Neural_Network_NN_with_1_75_trillion_parameters_the_start_of_Race_to_Quadrillion_parameters_NN
Dear Nima,
RL can be used in almost any application. It is a learning based on experience algorithm, a decision maker algorithm, an algorithm that learns autonomously, an optimization algorithm that over time learns to maximize it's reward, the reward can be defined by the engineer to reach the objective of the problem.
In mechanical engineering it can be used to control different kinds of robots including wheeled robot, ground robot, legged robot, UAV robot, etc.
The best book for RL in my opinion is 'Reinforcement Learning: An introduction', it has two versions, 1998, and 2018.
But if you want to apply RL and skip the theory you may start by reading the papers. In case of control and RL, you probably should use an RL algorithm that produces continuous action values such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO), I suggest the second one. Depending on the state of your problem, that is your sensors values, you might need to use shallow or deep neural network architectures as the function approximator of your RL.
For implementation of the RL algorithms the best is to refer to the OpenAI baseline repository in GitHub.
Good luck,
Amir Ramezani
Maybe the following open educMaybe the following open educational resource repository helps, which also comes along with learning videos and many programming examples from the engineering field:
https://github.com/upb-lea/reinforcement_learning_course_materials
- Badges
- Science topic