Training
Reward Prediction Error (RPE)
1997ActiveUpdated: 22 September 2026Published
Key
innovation
Formalizing 'reward surprise' as a learning signal shared by RL and neuroscience.
Category
Training
Abstraction level
Primitive
Operation level
Training
Use cases
Reinforcement learningBrain dopamine modelsInterpretability
How it works
Compute RPE = received reward − predicted reward; the signal updates predictions and the policy.
Problem solved
An agent must know how much the actual outcome deviates from expectations to update its behavior.