Surrogate
Use that as a “surrogate” - Kyle Vedder
A surrogate is:
- Any stand-in for a more complex process
- Used to approximate behavior, often for faster computation
- Common in black-box optimization, simulations, and reinforcement learning
How is this different from a Model Distillation?
Distillation is: A training procedure where a small model learns from a larger model. Typically supervised with soft labels (e.g. logits) from the teacher. Aims to preserve accuracy while shrinking model size.
“Unlike traditional supervised learning, where model selection can be achieved by using the model with the lowest validation loss [21],offline policy learning often suffers from the fact that the training objective is only a surrogate for the true objective of interest (e.g. task success rate)