waffledood/ME5406-Project2

A project to investigate different DRL methods for a self-driving car policy

Python

ME5406

Group 60

Ng Wei Jie, Brandon (A0184893L)
Mohamad Haikal Bin Mohamad Yusuf (A0182446B)
Dennis Goh Wen Qin (A0096927L)

Setup

Develop and test on Python3.6 in Ubuntu 18

sudo apt install python3-pip
python3 -m pip install --upgrade pip
python3 -m pip install virtualenv

Setup virtualenv

python3 -m virtualenv env
source env/bin/activate

Install python3 libraries

pip3 install -r requirements.txt

Train RL Models

Train DQN

is_eval=0 python3 main_dqn.py

Train DDQN

is_eval=0 python3 main_ddqn.py

Train A2C

is_eval=0 python3 main_a2c.py

NOTE: Rerun the program if the convergence is not as good as reported in the performance section

Evaluate RL Models

Evaluate DQN (model is stored as models/best_dqn.pt)

is_eval=1 python3 main_dqn.py

Evaluate DDQN (model is stored as models/best_ddqn.pt)

is_eval=1 python3 main_ddqn.py

Evaluate A2C (model is stored as models/best_a2c.pt)

is_eval=1 python3 main_a2c.py

PID Controller

Control with PID

python3 main_pid.py

Simulator

Control simulator with keyboard

python3 demo.py

Observation Space

- Game Engine

Performance

DQN

DDQN

A2C