Part 4 of an RL course series covering model-free learning methods. Topics include Monte Carlo prediction and control, temporal-difference learning, the bias-variance tradeoff between MC and DP, SARSA, Q-learning, and maximization bias. A hands-on experiment comparing SARSA vs. Q-learning on the Cliff Walking gridworld with code is included. No prior RL background required. The post also motivates RL's relevance to modern AI problems like LLM token generation, agentic pipelines, and RLHF/GRPO-based post-training.
Sort: