← AI Terminology
RL - Reinforcement Learning
Reinforcement learning is a machine learning paradigm where an agent learns to make decisions by interacting with an environment — receiving reward signals and adjusting its policy to maximise cumulative reward over time, without being given explicit labels for correct actions.
It is the third major ML paradigm alongside supervised and unsupervised learning.
It is the third major ML paradigm alongside supervised and unsupervised learning.
Why It Matters in AI
RL is the framework for sequential decision-making: it applies where supervised learning cannot, because there are no labels — only outcomes. AlphaGo/AlphaZero used RL to master Go beyond human ability; OpenAI Five used RL to defeat world champions at Dota 2; RLHF uses RL to align LLMs with human preferences. RL is also the mathematical foundation for robotics control, autonomous vehicles, and any system where the optimal action depends on long-term consequences.
Key Points
| Aspect | Description |
|---|---|
| Agent | The learner/decision-maker — the model being trained |
| Policy | π(a |
| Environment | What the agent interacts with — game, robot simulator, real world, or human rater |
| Exploration | ε-greedy, UCB, entropy bonus — agent must explore to discover better strategies |
| Reward signal | Scalar feedback after each action — the only training signal; designing it well is critical |
| Value function | V(s): expected cumulative reward from state s — learned to guide policy improvement |
Simple Analogy
Teaching a dog new tricks through rewards: no instructions, just a treat (positive reward) when the dog does the right thing and nothing (or a negative reward) when it doesn't. Over many trials, the dog discovers which behaviours earn treats and repeats them — that is RL, applied to an agent in an environment.
Common Usage Examples
gym.make("CartPole-v1")— classic RL environment for balancing a pole- PPO:
from stable_baselines3 import PPO; model = PPO("MlpPolicy", env); model.learn(100000) - AlphaGo: self-play RL — agent plays against itself, winning games are positive rewards
- RLHF:
PPOTrainer(model, ref_model, reward_model)— fine-tune LLMs via human preference rewards - Robotics: reward = distance to goal — RL agent learns locomotion policy from scratch in simulation
Summary
In short: Reinforcement learning trains agents to make sequences of decisions by maximising reward signals — the framework behind AlphaGo, robotics control, and RLHF alignment, applicable wherever optimal behaviour emerges from long-term consequences rather than labelled examples.