We provide different implementations of Deep Q-Learning working on multiple CPUs and one GPU for loss computation. All the implementations work for recurrent policies.
The implementations make use of the ReplayBuffer class which allows one to put a workspace and to get random workspaces. When putting a workspace, the workspace can be split into overlapping windows (i.e you can acquire trajectories of size 50, but store these trajectories as trajectories of size 2 in the replay buffer). Note that, since the ReplayBuffer returns a Workspace, it is very to replay an Agent on previously acquired trajectories just by dooing a forward pass.
- Double DQN: A Double DQN implementation with benchmark working with any trtypoe of policy (recurrent, non-recurrent, ...)
- Simple Double DQN: A simple implementation for non-recurrent policy used as a tutoria;