[Verse 1] An agent wanders through environments unknown Taking actions while the state machine's thrown Each decision yields a numerical prize Reward signals guide where wisdom lies Markov processes hold the golden key Future depends on present, not history [Chorus] RL framework spinning round and round Agent, environment, reward profound Policy maps the states to actions clear Bellman equations make the path appear Model-free or model-based you choose Q-learning wins when optimal's what you cruise [Verse 2] Value functions estimate the future gain Q-tables store what actions will obtain SARSA learns on-policy's steady beat While Q-learning off-policy can't be beat Deep networks handle spaces way too vast DQN bridges neural futures with the past [Chorus] RL framework spinning round and round Agent, environment, reward profound Policy maps the states to actions clear Bellman equations make the path appear Model-free or model-based you choose Q-learning wins when optimal's what you cruise [Bridge] Epsilon-greedy balances the scale Exploration versus exploit without fail UCB algorithms pull the bandit's arm REINFORCE gradients work their policy charm RLHF aligns the language models tight Human feedback keeps the outputs right [Verse 3] CartPole swinging in gymnasium's frame Training agents in the balance game Robotics, games, and recommendations flow Resource allocation makes the profits grow Temporal difference learning updates each step Neural networks keep the knowledge deeply kept [Chorus] RL framework spinning round and round Agent, environment, reward profound Policy maps the states to actions clear Bellman equations make the path appear Model-free or model-based you choose Q-learning wins when optimal's what you cruise [Outro] From bandits to the deepest neural maze Reinforcement learning lights tomorrow's ways
← Unit 2.2 — Unsupervised Learning | Unit 2.4 — ML Engineering Best Practices →