Tuning hyperparameters
In the previous section, we had Pong solved in three hours of optimization and 9M frames. Now it's a good time to tweak our hyperparameters to speed up convergence. The golden rule here is to tweak one option at a time and make conclusions carefully, as the whole process is stochastic.
In this section, we'll start with the original hyperparameters and perform the following experiments:
- Increasing the learning rate
- Increasing the entropy beta
- Changing the count of environments that we're using to gather experience
- Tweaking the size of the batch
Strictly speaking, the experiments below weren't proper hyperparameter tuning, just an attempt to get a better understanding of how A2C convergence dynamics depend on the parameters. To find the best set of parameters, the full grid search or random sampling of values could give much better results, but will require much more time and resources to conduct.
Learning rate
Our starting learning rate (LR) is 0.001...
