Implementation of Distributed Reinforcement Learning with Tensorflow

Information

• 20 actors with 1 learner.
• Tensorflow implementation with distributed tensorflow of server-client architecture.
• Recurrent Experience Replay in Distributed Reinforcement Learning is implemented in Breakout-Deterministic-v4 with POMDP(Observation not provided with 20% probability)

Dependency

opencv-python
gym[atari]
tensorboardX
tensorflow==1.14.0

Implementation

• [x] Asynchronous Methods for Deep Reinforcement Learning
• [x] IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
• [x] DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY
• [x] Recurrent Experience Replay in Distributed Reinforcement Learning

How to Run

• A3C: Asynchronous Methods for Deep Reinforcement Learning

CUDA_VISIBLE_DEVICES=-1 python train_a3c.py --job_name --job_name actor --task 0

CUDA_VISIBLE_DEVICES=-1 python train_a3c.py --job_name --job_name actor --task 0
CUDA_VISIBLE_DEVICES=-1 python train_a3c.py --job_name --job_name actor --task 1
CUDA_VISIBLE_DEVICES=-1 python train_a3c.py --job_name --job_name actor --task 2
...
CUDA_VISIBLE_DEVICES=-1 python train_a3c.py --job_name --job_name actor --task 19

• Ape-x: DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY

python train_apex.py --job_name learner --task 0

CUDA_VISIBLE_DEVICES=-1 python train_apex.py --job_name actor --task 0
CUDA_VISIBLE_DEVICES=-1 python train_apex.py --job_name actor --task 1
CUDA_VISIBLE_DEVICES=-1 python train_apex.py --job_name actor --task 2
...
CUDA_VISIBLE_DEVICES=-1 python train_apex.py --job_name actor --task 19

• IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures

python train_impala.py --job_name learner --task 0

CUDA_VISIBLE_DEVICES=-1 python train_impala.py --job_name actor --task 0
CUDA_VISIBLE_DEVICES=-1 python train_impala.py --job_name actor --task 1
CUDA_VISIBLE_DEVICES=-1 python train_impala.py --job_name actor --task 2
...
CUDA_VISIBLE_DEVICES=-1 python train_impala.py --job_name actor --task 19

• R2D2: Recurrent Experience Replay in Distributed Reinforcement Learning

python train_r2d2.py --job_name learner --task 0

CUDA_VISIBLE_DEVICES=-1 python train_r2d2.py --job_name actor --task 0
CUDA_VISIBLE_DEVICES=-1 python train_r2d2.py --job_name actor --task 1
CUDA_VISIBLE_DEVICES=-1 python train_r2d2.py --job_name actor --task 2
...
CUDA_VISIBLE_DEVICES=-1 python train_r2d2.py --job_name actor --task 39

Reference

1. IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
2. DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY
3. Recurrent Experience Replay in Distributed Reinforcement Learning
4. deepmind/scalable_agent
5. google-research/seed-rl
6. Asynchronous_Advatnage_Actor_Critic
7. Relational_Deep_Reinforcement_Learning
8. Deep Recurrent Q-Learning for Partially Observable MDPs