Multi-stage feature modeling and dynamic optimization strategy for time-series decision systems

Abstract

In the sequential decision-making scenarios of complex systems, the occurrence of key events is often influenced by the interaction of multiple features, with dynamic patterns characterized by nonlinearity, heterogeneity, and uncertainty. A significant challenge in intelligent decision-making research is how to achieve critical feature identification, state partitioning, and optimal strategy formulation under incomplete observation conditions. This paper presents a multi-stage modeling and optimization framework to address time-sensitive strategy selection problems. First, the distance correlation coefficient is used to select and structurally model multidimensional features to capture potential nonlinear dependencies. Second, the K-Means clustering algorithm is applied to automatically partition the state space, constructing a set of composite states with structural differences. On this basis, the semi-parametric Cox proportional hazards model is introduced to describe the dynamic distribution of event success probabilities over time under different states. Finally, the strategy selection is modeled as a dynamic programming problem over a finite time horizon, and the optimal decision path is solved using a backward iteration method based on the Bellman equation. Experimental results demonstrate that this framework effectively reduces decision risks, maintains stable performance under parameter perturbations, and exhibits strong robustness and scalability. This research provides a unified methodology for strategy optimization in time-sensitive systems, which can be widely applied to intelligent decision-making scenarios with uncertain states and complex features.