Slither-DRL

In this repository, we implement some well known deep reinforcement learning (DRL) algorithms for slither.io.

• Policy Gradient (PG)
• Deep-Q Network (DQN)
• Actor-Critic (AC)
• Advantage Actor-Critic (A2C)

Installation Instructions :

• Install docker for ubuntu 16.04 MAKE SURE TO DO STEP 2 AS WELL

• Install Conda for ubuntu 16.04

• Create Conda env

conda create --name slither python=3.5

• Activate a conda env

source activate slither

• Install needed packages

sudo apt-get update
sudo apt-get install -y tmux htop cmake golang libjpeg-dev libgtk2.0-0 ffmpeg

• Install pytorch 1.1.0 for CUDA 10.0

pip3 install https://download.pytorch.org/whl/cu100/torch-1.1.0-cp35-cp35m-linux_x86_64.whl
pip3 install https://download.pytorch.org/whl/cu100/torchvision-0.3.0-cp35-cp35m-linux_x86_64.whl

• Install universe installation dependencies

pip install numpy
pip install gym==0.9.5

• Install universe

git clone https://github.com/openai/universe.git
cd universe
pip install -e .

• Install this repository

cd ..
git clone https://github.com/JuiHsiu/Slither-DRL.git
cd Slither-DRL

How to Run :

• training the agent

python main.py --train_[pg|dqn|ac|a2c] 

• For DQN, there are some improvements can be added: (optional)

python main.py --train_dqn [--dueling_dqn] [--prioritized_dqn]

• testing the agent

python main.py --test_[pg|dqn|ac|a2c]

• If you want to see your agent playing the game,

python main.py --train_[pg|dqn|ac|a2c] --do_render

• By default, when you test the agent, the procedure is recorded as a video. You can assign the directory to the video by :

python main.py --test_[pg|dqn|ac|a2c] --video_dir [path_to_save_video]

Advanced Arguments :

1. Number of Environment

   You can create more than one environment at the same time. However, you need to modify the codes to perform batch learning.

python main.py --train_[pg|dqn|ac|a2c] --remotes [#_of_env]

2. Action Space

   We make 12 different positions of the mouse as the action space of our agent. If you want the agent to have the ability to accumulate, set the action_space = 24.

python main.py --train_[pg|dqn|ac|a2c] --action_space [12|24]

Demo :

Our best model is A2C and you can see the pre-trained agent playing game as following: