digshock / distributed_reinforcement_learning

Implementation of Distributed Reinforcement Learning with Tensorflow

Information

• 20 actors with 1 learner.
• Tensorflow implementation with distributed tensorflow of server-client architecture.

Dependency

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

Implementation

• IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
• DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY
• R2D2: Repeatable and Reliable Detector and Descriptor

How to Run

• 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

Reference

1. IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
2. DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY
3. R2D2: Repeatable and Reliable Detector and Descriptor
4. deepmind/scalable_agent
5. Asynchronous_Advatnage_Actor_Critic
6. Relational_Deep_Reinforcement_Learning