A Review of End-to-End Decision Optimization Research: An Architectural Perspective

Traditional decision optimization methods primarily focus on model construction and solution, leaving parameter estimation and inter-variable relationships to statistical research. The traditional approach divides problem-solving into two independent stages: predict first and then optimize. This decoupling leads to the propagation of prediction errors-even minor inaccuracies in predictions can be amplified into significant decision biases during the optimization phase. To tackle this issue, scholars have proposed end-to-end decision optimization methods, which integrate the prediction and decision-making stages into a unified framework. By doing so, these approaches effectively mitigate error propagation and enhance overall decision performance. From an architectural design perspective, this review focuses on categorizing end-to-end decision optimization methods based on how the prediction and decision modules are integrated. It classifies mainstream approaches into three typical paradigms: constructing closed-loop loss functions, building differentiable optimization layers, and parameterizing the representation of optimization problems. It also examines their implementation pathways leveraging deep learning technologies. The strengths and limitations of these paradigms essentially stem from the inherent trade-offs in their architectural designs. Through a systematic analysis of existing research, this paper identifies key challenges in three core areas: data, variable relationships, and gradient propagation. Among these, handling non-convexity and complex constraints is critical for model generalization, while quantifying decision-dependent endogenous uncertainty remains an indispensable challenge for practical deployment.