A reward scores what happened. It never states which action was correct, so the agent must infer the connection between its decisions and the outcome.
Two kinds of learning signal
Supervised label
- Supplies the correct output for each input
- The model knows exactly what it should have predicted
- Error is measured directly against the answer
Reinforcement reward
- Supplies a scalar score for the outcome
- The agent does not know which action earned the score
- Credit must be assigned across recent decisions
Delayed score in a game
An agent presses a button at time one and sees no immediate change. Three seconds later the score rises by one. The reward is real, but it is attached to a moment far removed from the action that caused it. If the agent had pressed several buttons in between, it cannot tell from the reward alone which press mattered. A supervised label would have said, at the moment of the press, whether that press was correct. The reward says only that the overall sequence ended well.
This gap between outcome and cause is called the credit assignment problem. It is why reinforcement learning needs repeated experience rather than a single labeled example: only many trials, with varying actions and outcomes, let the agent separate the decisions that help from those that do not.