Algorithms Implemented

• Q-Learning: Tabular approach to estimate state-action values and iteratively update Q-values based on rewards.
• Monte Carlo Methods: Episode-based learning with returns averaging to estimate state-action value functions.

AI Agent Development

• Designed and trained an AI agent to learn optimal blackjack strategies.
• State representation included hand value, dealer card, and usable ace conditions.
• Implemented reward functions reflecting wins, losses, and ties.

Training & Policy Optimization

• Simulated multiple episodes to converge on the optimal policy.
• Evaluated performance by comparing learned policy to known blackjack optimal strategies.
• Applied epsilon-greedy strategy for balancing exploration and exploitation.

Implementation Details

• Entire RL environment and agent implemented in JavaScript for browser-based simulation.
• Modular architecture separating environment logic, agent, and policy evaluation for maintainability.