https://giacolees.github.io/tags/policy-gradient/Recent content in Policy Gradient on giacolees - Tech BlogHugoenSat, 04 Apr 2026 12:14:26 +0200https://giacolees.github.io/posts/rl_intro/Sat, 04 Apr 2026 12:14:26 +0200https://giacolees.github.io/posts/rl_intro/TL;DR Reinforcement learning is trial-and-error learning: an agent takes actions in an environment to maximize cumulative reward. The math backbone is the Markov Decision Process and the Bellman equation, which expresses a state's value recursively through its successors. Value-based methods estimate how good states (or state-action pairs) are — planning with dynamic programming when the model is known, learning from experience with Monte Carlo and Temporal Difference updates when it isn't — and act greedily with respect to those estimates (SARSA on-policy, Q-learning off-policy).