Doom Reinforcement Learning Agent

A reinforcement learning agent developed on the VizDoom environment.

Setup

1. Install dependencies following the official documentation
2. pip install -r requirements.txt
3. Run python basic.py to test if installation was successful

How to control the script

The main script is agent.py, you can control what it runs using the constants.py. All the parameters are well described by their names, but {number} are most important:

• model_savefile: name of the file which model will be (was) saved;
• save_model: instruction to save model or just train and watch a new agent;
• load_model: instruction to load an existent model or not;
• visible_during_train: control if game image will appear during train or not (keep it False to speed up train process);
• show_labels: in future will control if agent use labels extracted with OpenCV.

If you want to train a new agent, set save_model = True and load_model = False. To run a existent model, set save_model = False and load_model = True (and choose the apropriate model name).

Observations for training

• According to the official documentation from ViZDoom, 5-10 epochs are okay to train a good model. I discovered that it is only true for some scenarios (e.g. basics modified scenarios), but I couldn't find a good number for all scenarios;
• Agents trained with a large number of epochs (50 or higher) start to decrease its learning capabilities for misterious reasons;
• Agents are saved in agents folder.
• After trained, if save_model = True, the script will generate a "Score x Epochs" graph on results folder.