Foundations
Reinforcement LearningRL
Also known as: learning from rewards
Training an AI by trial and error, rewarding good actions and penalising bad ones, so it learns a strategy that maximises reward.
Reinforcement learning is teaching by trial and error, like training a dog with treats. The AI tries something, gets a reward if it did well or nothing (or a penalty) if it did badly, and gradually learns which actions lead to the best outcomes. It is how AI learned to beat world champions at games like Go, and it helps make chatbots more helpful.
Reinforcement learning (RL) trains an agent to act in an environment so as to maximise a cumulative reward. The agent observes a situation, takes an action, receives a reward and a new situation, and over many attempts learns a policy, a strategy mapping situations to good actions.
What makes it different
Unlike supervised learning, there is no labelled "correct action" for each step. The agent must discover good behaviour through exploration, balancing trying new things against exploiting what already works, and often only gets feedback much later (a move that wins a game 50 turns on).
Where you see it
Game-playing AI (Go, chess, video games), robotics, and, importantly, aligning language models: RLHF uses RL to tune assistants toward human preferences, and reasoning models are trained with RL to think through problems.
Reinforcement learning formalises sequential decision-making as a Markov decision process, where an agent learns a policy π(a|s) to maximise expected discounted return via interaction, without explicit supervised targets (Sutton & Barto 2018). Core challenges are the exploration-exploitation trade-off, credit assignment over delayed rewards, and sample efficiency. Method families include value-based (Q-learning, DQN), policy-gradient/actor-critic (REINFORCE, PPO), and model-based approaches; deep RL couples these with neural-network function approximation, enabling landmark results such as AlphaGo (Silver et al. 2016).
In the LLM era RL is pivotal beyond games: RLHF and its variants optimise a policy against a learned human-preference reward for alignment, and RL against verifiable rewards (correctness on maths/code) trains reasoning models to produce long deliberate chains. Practical difficulties, reward misspecification/hacking, instability, and evaluation, mirror broader alignment concerns, since what you reward is what you get.