What Is Reinforcement Learning (RL)?

Three big families (and when to use each)

Value‑based methods (e.g., Q‑learning, DQN).
Learn how good it is to take an action in a state (the Q‑value), then act greedily with respect to value. Great when actions are discrete and you can explore a lot.

Policy‑gradient methods (e.g., REINFORCE, PPO).
Learn the policy directly by nudging its parameters to increase expected reward. These shine with continuous actions (robotics, control) and when you can simulate experiences.

Model‑based RL.
Learn a model of the environment to plan ahead. Sample‑efficient and powerful when you can approximate dynamics (e.g., industrial processes) or mix real with simulated data.

Modern systems blend these ideas: value baselines inside policy‑gradient methods, learned world models for imagination‑based planning, and offline RL to leverage existing logs.

Why RL is hot again

• Agents, not just predictors. As LLMs gain tool‑use, RL provides a principled way to optimize behavior (weighing delayed effects) instead of just optimizing next‑token probabilities.
• Real‑world wins. From robotics coordination to process optimization, RL is moving from research to production.
• Feedback beyond labels. You can encode business objectives as rewards (SLA compliance, energy use, risk limits) and let the system learn trade‑offs.

Where RL shines (and where it doesn’t)

Great fits: robotics manipulation and coordination; dynamic pricing and ad allocation; supply chain control; energy optimization; agentic planning for LLM tools; A/B‑tested UX policies.
Tough fits: scarce feedback, high‑stakes safety without sandboxing, non‑stationary rewards, and situations where a rule‑based policy already nails it.

‍