fix(pytorch): correct step-based learning rate decay

rickstaa / stable-learning-control

A framework for training theoretically stable (and robust) Reinforcement Learning control algorithms.

https://rickstaa.dev/stable-learning-control

MIT License

6 stars 1 forks source link

fix(pytorch): correct step-based learning rate decay #405

Closed rickstaa closed 7 months ago

rickstaa commented 7 months ago

This pull request addresses an issue with the step-based learning rate decay mechanism when lr_decay_ref is set to 'step'. Previously, the learning rate was decaying too rapidly due to a bug in the decay logic. This fix ensures that the learning rate decays at the correct pace as per the step-based decay configuration.