ingeechart/DAgger_example

DAgger_example

tested on ubuntu 18.04LTS

Installation

1. install torcs

   cd vtorcs-RL-color
   sudo apt-get install libglib2.0-dev  libgl1-mesa-dev libglu1-mesa-dev  freeglut3-dev ibplib-dev  libopenal-dev libalut-dev libxi-dev libxmu-dev libxrender-dev  libxrandr-dev   libpng-dev
   ./configure
   sudo make
   sudo make install 
   sudo make datainstall
   

2. install xautomation for key automation

   sudo apt-get update -y
   sudo apt-get install -y xautomation
   

Training Learner

first need to change configureation of game environment

# start torcs env
sudo torcs -vision

1. go to race configuration and select racing track and set the driver scr_server 1
   • race -> practice -> configure race -> select racing track -> select driver
2. go to display setup, select 64x64 RGB observation and 16 color depth
   • options -> Display

start training!

python train.py

Structure

DAgger_example
├── README.md
├── train.py
│     train agent with DAgger algorithm
│
├── agent.py 
│     neural network architecture of learner
│
├── snakeoil3_gym.py
│     Torcs client communicate with Torcs game server
│ 
├── gym_torcs.py
│     make Torcs client gym environment
│
├── autostart.sh
│     key automation
│
├── vtorcs-RL-color
│    Torcs source, copied from https://github.com/ugo-nama-kun/gym_torcs
│
├── images
        sampled images