핵심 개념
Input Convex Loss Network (ICLN) proposes a global surrogate loss model for Decision Focused Learning, learning a mapping from prediction to task loss.
초록
ICLN introduces a novel global surrogate loss model for Decision Focused Learning, addressing the challenge of integrating prediction and optimization paradigms. The content discusses the architecture of ICLN, algorithms for training ICLN-L and ICLN-G, and experiment details on the Inventory Stock Problem.
Abstract
Decision-focused Learning (DFL) integrates prediction and optimization.
Existing methods use surrogate optimization or loss functions.
ICLN proposes a global surrogate loss model for DFL.
Introduction
Decision-making under uncertainty involves prediction and optimization tasks.
Prediction-Focused Learning (PFL) and Decision-Focused Learning (DFL) differ in their approach.
DFL aims to train models for good decisions directly.
Input Convex Loss Network
ICLN learns task loss via Input Convex Neural Networks.
ICLN-L represents local task loss, while ICLN-G represents global task loss.
Related Works
Various methodologies based on gradient-based learning in DFL.
Surrogate DFL models aim to obtain gradients effectively.
Experiments and Results
Three experiments conducted to evaluate ICLN-L and ICLN-G.
ICLN-G shows promising results with significantly fewer training samples.
Conclusion
ICLN offers a handcraft-free approach to learning surrogate loss models.
ICLN-G reduces time complexity for training ML models in DFL pipelines.
통계
ICLN은 Decision Focused Learning을 위한 전역 대리 손실 모델을 제안합니다.
인용구
"ICLN proposes a global surrogate loss model for DFL."
"ICLN-G shows promising results with significantly fewer training samples."