: Shoot randomly generated targets on the opposite wall with a gun.

: 3 actions [Left, Right, Shot]

-1 if 1 frame has passed -5 if shot missed +100 if shot hit

(1) 200 frame timeout (2) enemy hit

Warning: You might have to tune some file paths in the code.

basic.py
Run to test basic random policy agent in ViZDoom environment.

no_trained_base_model.py
Run to test initialized CNN-based agent.

trained_base_model.py
You can make model trained with ViZDoom scenario by REINFORCE policy gradient.
Reward and Loss will be saved during training every episodes.
Your trained model will be saved every after 10000 episodes.

gradient_accumulation.py
This is a version with gradient accumulation algorithm added to the trained_base_model.py.
Except for that, everything else is the same.

play_model.py
You can see how your trained agent play ViZDoom scenario well via this file.
Note that this file has to be excuted in terminal.
How to excute this file is writed at the last line of the file.

graph_log.py
Produces graph.png file via log.txt or acc_log.txt.
Tune some variables to tune value for x or y axis.

model.py
CNN model is factorized in this file.

for_ppt.py
It's just a dump file.

Jeonghyun Kim 📧 [email protected]

Jiyoung Seo 📧 [email protected]