Training
Temporal Difference error (TD error)
1988ActiveUpdated: 22 September 2026Published
Key
innovation
Learning from bootstrapping — correcting predictions from the next estimate, step by step.
Category
Training
Abstraction level
Primitive
Operation level
Training
Use cases
Value-function learningQ-learning / SARSAActor-critic
How it works
Compute δ = r + γV(s′) − V(s) and update V(s) toward reducing this error.
Problem solved
We need to learn state values without waiting for episode end and without a full model of the environment.