Why Reinforcement Learning Matters
Before defining reinforcement learning formally, start with a stranger question: what is learning? Useful answers include feedback, compression, pattern finding, exposure, repetition, and the ability to use information to improve future behavior.
Reinforcement learning is easiest to misunderstand when it gets framed as another training loss. RL is learning through experience: an agent acts, receives consequences, and changes future behavior.
Learning in the brain
Case studies before the RL loop
Each example highlights a different premise: experience changes representations, learning systems can separate, and practice can produce task-linked structural change.
London taxi drivers
A learner's experience stream can reshape the internal map it uses to act.
Licensed London taxi drivers spend years mastering a dense city layout. MRI studies found navigation-related hippocampal differences, especially in posterior regions associated with spatial representation.
Why it matters
Experience can change the representation that future decisions depend on.
London taxi drivers
Learning as feedback
Trace a path with each hand
Use your normal drawing hand for one attempt, then switch to your opposite hand. Repeat and watch how feedback plus practice changes performance.
Choose a hand, then trace the gray path.
Best usual
-
Best opposite
-
Attempts
0
Attempt history
Trace with each hand a few times. The first opposite-hand attempt is usually the interesting one.
Learning leaves traces in the learner. London taxi drivers, patient H.M., and juggling studies all point toward mechanistic changes or separable memory systems. Practice needs the right feedback, memory, and task structure to become improvement.
That matters for RL because an agent also changes through experience. The experience stream depends on the agent's current behavior. A poor policy can generate poor data; a better policy can discover better data.
The Core Difference
Supervised learning solves prediction problems with a fixed dataset and labels. Reinforcement learning solves decision problems where the agent's actions help determine what data it sees next.
Why is RL data policy-dependent?