Ape-X variations of DQN and DDPG (`APEX_DQN <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/dqn/apex.py>`__, `APEX_DDPG <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/ddpg/apex.py>`__ in RLlib) use a single GPU learner and many CPU workers for experience collection. Experience collection can scale to hundreds of CPU workers due to the distributed prioritization of experience prior to storage in replay buffers.
RLlib's A3C uses the AsyncGradientsOptimizer to apply gradients computed remotely on policy evaluation actors. It scales to up to 16-32 worker processes, depending on the environment. Both a TensorFlow (LSTM), and PyTorch version are available. Note that if you have a GPU, `IMPALA <#importance-weighted-actor-learner-architecture>`__ probably will perform better than A3C.
DDPG is implemented similarly to DQN (below). The algorithm can be scaled by increasing the number of workers, switching to AsyncGradientsOptimizer, or using Ape-X.
RLlib DQN is implemented using the SyncReplayOptimizer. The algorithm can be scaled by increasing the number of workers, using the AsyncGradientsOptimizer for async DQN, or using Ape-X. Memory usage is reduced by compressing samples in the replay buffer with LZ4.
In IMPALA, a central learner runs SGD in a tight loop while asynchronously pulling sample batches from many actor processes. RLlib's IMPALA implementation uses DeepMind's reference `V-trace code <https://github.com/deepmind/scalable_agent/blob/master/vtrace.py>`__. Note that we do not provide a deep residual network out of the box, but one can be plugged in as a `custom model <rllib-models.html#custom-models>`__.
RLlib's IMPALA implementation scales from 16 to 128 workers on PongNoFrameskip-v4. With vectorization, similar learning performance to 128 workers can be achieved with only 32 workers. This is about an order of magnitude faster than A3C (not shown here), with similar sample efficiency.
`[paper] <https://papers.nips.cc/paper/1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf>`__`[implementation] <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/pg/pg.py>`__ We include a vanilla policy gradients implementation as an example algorithm. This is usually outperformed by PPO.
PPO's clipped objective supports multiple SGD passes over the same batch of experiences. RLlib's multi-GPU optimizer pins that data in GPU memory to avoid unnecessary transfers from host memory, substantially improving performance over a naive implementation. RLlib's PPO scales out using multiple workers for experience collection, and also with multiple GPUs for SGD.
