Add a dropout layer

[sallem]

Hey, I would like to add dropout layers into mu DQN model.

A dropout is a layer useful for training a more robust model by randomly removing some values during the training in order to generalize it.

Is it possible to add it with the config file ? Did not find this information into the documentation.

[karroyan]

Hi, thank you for raising this issue.

We plan to implement the ability to add layers via the config file within this week. We will open a pull request once this feature is ready. Please keep an eye on the repository for updates.

[karroyan]

@sallem, hi, in which environments you plan to use DQN?

[sallem]

Hey

I am using it with stocks environment for trading.

[karroyan]

Hi, the dropout control has been implemented in pull request #712 . You can now add a dropout=0.3 parameter under the config.policy.model configuration to enable dropout regularization with dropout probability of 0.3 after each linear layer followed by activation in DQN with FCEncoder. The configuration example can be found in cartpole_dqn_config. Let me know if you have any other questions!