Supervised, Unsupervised, and Reinforcement Learning

Machine learning is not one technique but a family of approaches. The three main paradigms — supervised, unsupervised, and reinforcement learning — describe how a model receives signal about what "good" looks like. Understanding these categories helps you interpret model behavior and evaluate outputs in context.

Supervised Learning

The model learns from labeled examples: inputs paired with correct outputs. A spam filter trained on emails marked "spam" or "not spam." An image classifier trained on photos tagged with object names. A text model fine-tuned on prompt-response pairs written by humans.

Supervised learning is the most common paradigm in production AI. When you rate whether a chatbot response is accurate, you are providing the kind of labeled signal that supervised systems learn from.

Evaluation focus: Does the output match what a correct label would look like? Are labels consistent across similar examples?

Unsupervised Learning

The model finds structure in unlabeled data without being told the right answers. Clustering similar customers, detecting anomalous transactions, or learning compressed representations of text. The model discovers patterns on its own.

Unsupervised methods are often used for data exploration, preprocessing, or as a step before supervised fine-tuning.

Evaluation focus: Are the discovered patterns meaningful? Do clusters or anomalies align with real-world categories?

Reinforcement Learning

An agent takes actions in an environment and receives rewards or penalties based on outcomes. The agent learns a policy that maximizes cumulative reward. Classic applications include game-playing AI, robotics, and recommendation systems that optimize for engagement.

Reinforcement Learning from Human Feedback (RLHF) adapts this paradigm for language models: humans rate outputs, a reward model learns those preferences, and the language model is tuned to produce higher-rated responses.

Evaluation focus: Does the behavior align with intended goals? Are there reward-hacking shortcuts where the model optimizes the metric but not the underlying objective?

How They Combine

Modern AI products layer all three. A language model might be pre-trained unsupervised on web text, fine-tuned supervised on curated examples, and aligned via RLHF with human ratings. Your evaluation tasks may touch any of these stages.

Choosing the Right Mental Model

When a task instructions mention "preference ranking," think reinforcement learning. When you label individual responses as correct or incorrect, think supervised learning. When you assess whether clusters or groupings make sense, you may be evaluating unsupervised output. Matching your evaluation mindset to the underlying learning paradigm helps you provide feedback the engineering team can actually use.

Key Takeaways

• Supervised learning uses labeled examples; most evaluation tasks assume this paradigm
• Unsupervised learning discovers patterns without labels
• Reinforcement learning optimizes behavior through rewards — including human feedback
• Production systems often combine all three approaches