学习未知泊松过程下的最优停车策略，实现近最优搜索决策。
Learning Optimal Search Strategies
- 通过估计累积跳跃强度而非强度函数来学习阈值策略
- 在广泛环境中实现对数级后悔增长，逼近理论最优
- 适合研究在线决策与自适应搜索的学者参考
我们研究了在停车问题中如何学习最优搜索策略，其中停车机会以未知非齐次泊松过程出现。最优策略为基于无差异位置的阈值型停止规则。本文提出一种算法，通过估计累积跳跃强度而非强度函数本身来学习该阈值。我们证明该算法在一大类环境上具有对数级后悔增长。此外，我们建立了对数级最小最大后悔下界，证实了所提方法的增长最优性。
原文摘要 · Abstract (English)
We explore the question of how to learn an optimal search strategy within the example of a parking problem where parking opportunities arrive according to an unknown inhomogeneous Poisson process. The optimal policy is a threshold-type stopping rule characterized by an indifference position. We propose an algorithm that learns this threshold by estimating the integrated jump intensity rather than the intensity function itself. We show that our algorithm achieves a logarithmic regret growth, uniformly over a broad class of environments. Moreover, we prove a logarithmic minimax regret lower bound, establishing the growth optimality of the proposed approach.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。