Reinforcement Learning: Cliff Walking

Overview

This repository contains the implementation of two fundamental reinforcement learning algorithms, Q-learning and SARSA, applied to the Cliff Walking environment. The project explores how these algorithms learn to navigate the gridworld, avoid the cliff, and reach the goal while minimizing penalties.

Key Concepts

• Parts:

  • Q-learning: An off-policy algorithm that learns the optimal policy by estimating the maximum future rewards.
  • SARSA: An on-policy algorithm that updates its policy based on the actual actions taken, leading to potentially safer but less aggressive strategies.
  • Comparison: A detailed comparison of the paths chosen by each algorithm, highlighting differences in exploration and exploitation behaviors.

• Tasks:

  • Implement and evaluate the Q-learning algorithm.
  • Implement and evaluate the SARSA algorithm.
  • Compare and analyze the optimal policies derived from both algorithms.