learned with Expected SARSA control.
This repository contains the code for the final project of A.Y. 2022-2023 Professor Antonio Celani's course on Reinforcement Learning at ICTP (International Centre for Theoretical Physics at Trieste, Italy).
This repository contains a study on the application of Temporal Difference learning methods to solve the Racetrack problem as described in the book "Reinforcement Learning: An Introduction" by Richard S. Sutton and Andrew G. Barto. The Racetrack problem is a classic Reinforcement Learning problem where an agent must navigate a racecar around a racetrack as efficiently as possible.
The primary focus of this study is to implement and compare different TD learning algorithms, including SARSA, Q-Learning, and Expected SARSA. The goal is to analyze how these algorithms perform in this specific environment, their convergence, and the quality of the learned policies.
This repository implements the following TD learning algorithms for solving the Racetrack problem:
-
SARSA (State-Action-Reward-State-Action): A TD control algorithm that estimates the Q-values for each state-action pair and learns a policy that chooses actions based on these estimates.
-
Q-Learning: Another TD control algorithm that estimates Q-values, but it learns an optimal policy by always choosing the action with the maximum Q-value in a given state.
-
Expected SARSA: A variant of SARSA that calculates the expected value of the next Q-value, taking into account the probability of taking different actions.
This project is licensed under the MIT License - see the LICENSE file for details.