jdhruva03/deep-rl---carracing

Using Deep Reinforcement Learning to Solve OpenAI’s ‘CarRacing’ Environment

Undergraduate Thesis

Using Deep Reinforcement Learning to Solve OpenAI’s ‘CarRacing’ Environment

Reinforcement learning is a subset of machine learning which involves an agent learning from interactions with an environment. We maximize the agent’s reward and through that learn good behavior for our agent. Reinforcement learning algorithms have proven effective in solving sequential decision-making problems in various fields.OpenAI Gym is a popular platform for testing and benchmarking reinforcement learning algorithms, offering different environments for testing agents. One such environment is the racing gym, which requires the agent (a car) to navigate a racing track, avoid obstacles and complete laps as quickly as possible. In this project, we try to solve OpenAI’s racing gym. The text explores two popular reinforcement learning algorithms, Deep Q learning (DQN) and Proximal Policy Optimization (PPO), for solving the racing gym problem. The text compares the performance of the two algorithms, investigates various hyperparameters, and identifies the optimal hyperparameters to train the final models. We discuss some key observations that went into setting up and performing the experiments for hyperparameter tuning. The text can be used as a reference by others to solve similar environments, and save time doing so, using the described reinforcement learning methods in this project.