search

Kaixhin / Rainbow

Rainbow: Combining Improvements in Deep Reinforcement Learning
MIT License
1.56k stars 282 forks source link

Hyper parameter #56

Closed jiafei1224 closed 4 years ago

jiafei1224 commented 4 years ago

Can i check with you on the difference between the canonical parameter and the data efficient parameter? Can i say that the data efficient parameter are more likely recommended?

Kaixhin commented 4 years ago

You can refer to the data-efficient Rainbow paper for details, but the hyperparameters were tuned there to achieve high rewards quickly, which doesn't necessarily mean that the performance after a lot of experience will be as good as they were with the original hyperparameters. But if you want to be "data-efficient", then yes the new hyperparameters should be a better starting point.