Training
Value function
1957ActiveUpdated: 22 September 2026Published
Key
innovation
Reducing a long-term goal to a value of states/actions that can be estimated and optimized.
Category
Training
Abstraction level
Primitive
Operation level
TrainingInference
Use cases
Q-learning / DQNActor-criticRLHF for LLMs
How it works
V(s)/Q(s,a) estimate expected discounted return; they are learned via TD or Monte Carlo and used to choose actions.
Problem solved
An agent needs an estimate of the long-term value of states/actions, not just the immediate reward.