3 threads
after 4000 episodes, about 30% chance of Clearance
after 8000 episodes, about 70% chance of Clearance
Install NES emulator FCEUX
lua scripts and nes files in 'SuperMarioBros' Folder are modified from gym-super-mario
Install pywin32
using win32 named pipe to comunicate with FCEUX, so support winodws only
config.py
FCEUX_PATH = 'D:\\fceux-2.2.3-win32\\fceux.exe'
change to your install path
A3CTrainer.py
load_model = False
change to True if you need to continue training
Train
python A3CTrainer.py
Test
python A3CTester.py
introduction article to this impelement
A3C model implements is from this article 《simple-reinforcement-learning-with-tensorflow-part-8-asynchronous-actor-critic-agents-a3c》.
I have tried a lot of A3C implements and this one is most efficient.