Miffyli / rl-human-prior-tricks

Evaluating different engineering tricks that make RL work

Reinforcement Learning Tricks, Index

This repository contains the code for the paper "Distilling Reinforcement Learning Tricks for Video Games".

Short story shorter: RL algorithms are neat and all, but to get it to work in video games (RL competitions and whatnot), there are some nifty little tricks involved that need bit of expertise in the domain. This includes reward shaping, curriculum learning, splitting task into subtasks by hand and guiding agent's actions. We took some of these tricks and tried them on three environments with DQN. With right setup you get more out of DQN.

Code authors: Anssi Kanervisto, Christian Scheller and Yanick Schraner.

The experiments in the three environments are split into three git branches:

• vizdoom for ViZDoom Deathmatch experiments
• minerl for MineRL ObtainDiamond experiments
• gfootball for Football environment experiments

To run the experiments, checkout the repository you want to run experiments for with git checkout [branch name], and follow the instructions in the README file there.

After running all the experiments, collect the results as described the respective branches. You should have three directories

• vizdoom-runs
• minerl-runs
• football-runs

After this, running python plot_paper.py should create a figures/learning_curves.pdf file which summarizes the results.