In the process of reinforcement learning training, the training was terminated erroneously because the data dimensions did not match.

I am trying to transfer the "Training with Reinforcement Learning" from the sample program "Train a Humanoid Walker" to another humanoid robot.But every time the training reaches the 18th episode, it will terminate with an error.

I don't know why there is no problem with the dimension of the first few episodes, but there is a problem with the 18th episode.

I tried to set "trainingOptions.StopOnError" to "off", but it didn't work. Although the program will not jump out of the error, it still stuck in the error every episode and could not continue training.

How should I eliminate this error? Thank you.

Answer 1

[Emmanouil Tzorakoleftherakis]

Are you using GPU to train? If yes, there is an indentified bug in R2020a which will be fixed when the first update for this release goes live. See post below:

https://www.mathworks.com/matlabcentral/answers/515391-error-using-gpu-in-matlab-2020a-for-reinforcement-learning

[Jingchao Li]

Yes，I do used GPU. Thank you!

Answer 2

[Jingchao Li]

When I go back to the 2019b version to run, no error is reported