Objective function
The quantity training tries to optimize — usually minimizing loss, sometimes maximizing reward.
Think of it like
Like the scoreboard a whole team plays toward — everything they do is in service of moving that one number the right way.
Example
In supervised training the objective is to minimize loss; in RL it’s to maximize expected reward.
How it actually works
The objective function is the general term for what optimization targets; a loss function is the common minimize-this special case, while reinforcement learning maximizes an objective like expected return. It also often bundles in regularization terms that trade raw fit against simplicity. Whatever it is, the model will optimize it literally — which is why mis-specified objectives are a core source of surprising, unwanted behavior.
For product teams
The single thing your model is truly trying to do; if it’s wrong, no amount of data fixes it.
For engineers
The optimized quantity (loss to minimize or reward to maximize), possibly with regularizers; literally optimized, so specify carefully.
Related
- Loss function — A minimize-case is the loss function.
- Reward model — In RL it’s the reward from the reward model.
- Regularization — May include regularization.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome