This project implements and trains sample-based reinforcement learning (RL) algorithms from scratch in Python, using the Taxi-v3 environment from OpenAI's Gym. The focus is on analyzing the performance of different algorithms, including Monte Carlo methods and Temporal Difference (TD) methods such as Q-learning and SARSA.
- Updates occur only at the end of each episode.
- Evaluates the agent’s performance and convergence speed over multiple episodes.
- Updates occur at every step, leading to faster convergence compared to Monte Carlo methods.
- Compares Q-learning and SARSA algorithms to evaluate variance in rewards and training speed.
- TD methods generally exhibited lower variance in rewards and faster training speeds compared to the Monte Carlo method.
- Among Q-learning and SARSA, Q-learning demonstrated less variance in rewards and lower error percentages.
The performance metrics and analysis of the implemented methods are documented in the provided report. Please refer to the respective files for comprehensive insights into the training outcomes.