Implementation of IMPALA with Distributed Tensorflow

Information

• These results are from only 32 threads.
• A total of 32 CPUs were used, 4 environments were configured for each game type, and a total of 8 games were learned.
• Tensorflow Implementation
• Use DQN model to inference action
• Use distributed tensorflow to implement Actor
• Training with 1 day
• Same parameter of paper

start learning rate = 0.0006
end learning rate = 0
learning frame = 1e6
gradient clip norm = 40
trajectory = 20
batch size = 32
reward clipping = -1 ~ 1

Dependency

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

Overall Schema

Model Architecture

How to Run

• show start.sh
• Learning 8 types of games at a time, one of which uses 4 environments.

Result

Video

Breakout	Pong	Seaquest	Space-Invader
Boxing	Star-Gunner	Kung-Fu	Demon

Plotting

Compare reward clipping method

Video

abs_one	soft_asymmetric

Plotting

abs_one
soft_asymmetric

Is Attention Really Working?

• Above Blocks are ignored.
• Ball and Bar are attentioned.
• Empty space are attentioned because of less trained.

Todo

• Only CPU Training method
• Distributed tensorflow
• Model fix for preventing collapsed
• Reward Clipping Experiment
• Parameter copying from global learner
• Add Relational Reinforcement Learning
• Add Action information to Model
• Multi Task Learning
• Add Recurrent Model
• Training on GPU, Inference on CPU

Reference

1. IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
2. deepmind/scalable_agent
3. Asynchronous_Advatnage_Actor_Critic