Function Approximation in RL

Part 5 of an RL series covering function approximation — the technique needed when tabular methods break down for real-world, continuous-state problems. Topics include why lookup tables fail, parameterized function approximators, gradient Monte Carlo, semi-gradient TD, and the deadly triad of function approximation with bootstrapping and off-policy learning. Includes a hands-on implementation training an agent on the Mountain Car problem. The post also contextualizes RL's growing importance given its role in post-training pipelines for LLMs like DeepSeek-R1, ChatGPT, and Claude.