Modify weights update rule of DDPG traning

Currently, i have studied DDPG algorithm for my control project. I want to modify a bit in the weight update rule (gradient descent) to reduce the steady state error of system. How can i do this?

Answer 1

[Gojo]

Hey Vu,

In order to make modifications in your gradient descent function, you can make changes in "rlOptimizerOptions" object as mentioned in the documentation for DDPG agent: https://www.mathworks.com/help/releases/R2023b/reinforcement-learning/ug/ddpg-agents.html

In order to make changes in the weight update rule,

• modify the "LearnRate" to change the rate at which the the model is converged.
• modify the "L2RegularizationFactor" to avoid overfitting.

Have a look at the "rlOptimizerOptions" for more details:

https://www.mathworks.com/help/reinforcement-learning/ref/rl.option.rloptimizeroptions.html#d126e67296

Happy coding!

[Vu]

Thank you for your comment. I have tried modify the 'LearnRate' but it seem not work, the steady state error still available. I want to add an Integral component to weight update rule as some papers suggest. Is there any way to do it?