Skip to content
VizDoom Bot Trained With Deep Q Reinforcement Learning
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
bashScripts
models
previousRuns
researchpapers
scenarios
Doom2.wad
README.md
_vizdoom.ini
pytorch.yaml
vanDamme.py

README.md

VanDamme

Arnold is a PyTorch implementation of a VizDoom bot trained using Deep Q Reinforcement Learning.

This repository contains:

  • The source code to train DOOM agents
  • Scenarios for the bot to perform on
  • Pretrained Models
  • Bash Scripts to load and train batches of models at once

Installation

Dependencies

VanDamme was written and tested on a Linux machine and all dependencies can be installed using the Anaconda environment yaml file: pytorch.yaml

Code structure

.
├── bashScripts     # Bash Scripts to Load and Save Multiple Models
├── models          # Pretrained Models
├── previousRuns    # Text files of full output of previous runs
├── researchpapers  # Example Research Papers on vizdoom training and DQN
├── scenarios       # Scenarios and wad files to load
├── vanDamme.py     # Main File
├── pytorch.yaml    # Anaconda yaml file
├── Doom2.wad    	# Assets from original Doom 2, all models were trained using this   
└── README.md

Using the VanDamme Parser

The parameters that can be specified during a run of VanDamme are shown below:

python vanDamme.py

## Arguments used by the vanDamme parser and their default settings
--run_name "ExampleName"        
--learning_rate .00025     
--epochs 20                   
--learning_steps_per_epoch 2000 
--replay_memory_size 10000  
--model_savefile "unnamedModel.pth"
--save_model False
--load_model False
--scenario_config "basic.cfg"
--set_to_shaping "False"
--doom2_wad "Doom2.wad"
You can’t perform that action at this time.