Arizona State University, Arizona

Ira A. Fulton Schools of Engineering

---

EEE 598: Reinforcement Learning in Robobotics Final Project

Reinforcement Learning for Self-Driving Cars

---

• This project is a Final Project carried out by Aaryan Bhardwaj, Apoorv Uplal, Aditya Ravichander, Akshat Sharma as part of our EEE 598 curriculum.
• Download here: Final-Report.pdf

---

Abstract

In this project, we implement a supervisor agent that uses deep reinforcement learning to improve driver assistance systems. The agent has to maximize its speed on a multi lane expressway. Collision avoidance function is added to the environment in order to avoid potentially dangerous scenarios. Additionally, the highway driving instance is a stochastic environment with built-in randomness. In this project, a traffic environment simulation, a neural network model, and a reinforcement learning agent is developed. The suggested model just makes use of the barest amount of environmental sensory data. The ideal policy is effectively learned by the agent and applied. To evaluate the model's statistical performance, three different traffic scenarios were investigated.

This work uses Songyan Ho's work as its starting point and focuses on observing variations in model behaviour with changes in reward structures.

Simulator

Simulator running under macOS environment

Run

1. Install latest Python version
2. Open Terminal. Navigate to main directory.
3. Install requirements using
      pip install -r requirements.txt
4. For training the RL agent, Open config.py, comment the lines 21, 22 & 23 and uncomment the lines 16, 17 & 18. Set the penalties and rewards in config.py
5. For setting the reward for overtaking and penalty for being overtaken, open gui_util.py and edit the lines 223 (Overtaking reward) and 227 (being overtaken penalty)
6. Run using command
      python gui.py
7. For testing the RL agent, Open config.py, uncomment the lines 21, 22 & 23 and comment the lines 16, 17 & 18.
8. Run using command
      python gui.py