یک چارچوب الگوریتم ریاضی برای راه حل های اکتشافی برای برنامه های پویای تصادفی افقی

Rollout algorithms have enjoyed success across a variety of domains as heuristic solution procedures for stochastic dynamic programs (SDPs). However, because most rollout implementations are closely tied to specific problems, the visibility of advances in rollout methods is limited, thereby making it difficult for researchers in other fields to extract general procedures and apply them to different areas. We present a rollout algorithm framework to make recent advances in rollout methods more accessible to researchers seeking heuristic policies for large-scale, finite-horizon SDPs. We formalize rollout variants exploiting the pre- and post-decision state variables as a means of overcoming computational limitations imposed by large state and action spaces. We present a unified analytical discussion, generalizing results from the literature and introducing new results that relate the performance of the rollout variants to one another. Relative to the literature, our policy-based approach to presenting and proving results makes a closer connection to the underpinnings of dynamic programming. Finally, we illustrate our framework and analytical results via application to a dynamic and stochastic multi-compartment knapsack problem.