Ding Sirui
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ding Sirui has a total of 3 co-authors. Here are the top co-authors by number of shared publications:
- Xiwen Bai
Tsinghua University
2 shared publications - Liangqi Cheng1 shared publication
- Xi Lin
Department of Civil and Environmental Engineering, University of Michigan, Ann Arbor, Michigan 48109
1 shared publication
Latest Publications
- Open Access
- 1 August 2026
- Published by Elsevier in SSRN Electronic Journal
Post-Decision Optimization-Decoder Reinforcement Learning for Maritime Fleet Renewal under Regime-Switching Market Uncertainty
Maritime fleet renewal requires long-lived investment and contracting under cyclical market uncertainty. This paper develops an integrated artificial intelligence-operations research (AI--OR) framework linking data-driven market uncertainty generation with efficient long-horizon decision optimization and evaluates it through a Very Large Crude Carrier case study. For uncertainty generation, a Gaussian hidden Markov model fitted to historical freight-market index data identifies latent regimes and transitions, while regime-conditional bootstrapping samples market scenarios. Compared with stylized processes or independent sampling, this design preserves market-cycle dynamics and empirical co-movement among freight rates, vessel prices, and market availability limits. The generated scenarios provide the market environment for fleet renewal decisions. For renewal decision making, existing studies often rely on multi-stage stochastic programming solved by rolling-horizon sample average approximation (RH-SAA). However, repeated finite-horizon optimization is computationally costly and may undervalue long-lived actions with short look-ahead windows. We reformulate the problem as a finite-horizon Markov decision process with a strategic-operational decomposition and develop the Post-Decision Optimization-Decoder Reinforcement Learning (POD-RL) framework. POD-RL learns state-dependent latent valuations for post-decision fleet and contract resources, while an optimization decoder converts them into feasible high-dimensional strategic actions. In two 20-year tests under homogeneous and nonhomogeneous state transitions, POD-RL improves mean cumulative profit over RH-SAA by 16.3% and 19.1%, respectively, while reducing online decision time from over 7,000 s to 0.1-0.2 s. When deployed without retraining in a 1999-2018 historical backtest, POD-RL achieves 18.8% higher realized profit than RH-SAA. The results indicate that the integrated framework improves long-horizon decision quality.
- 1 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPost-Decision Optimization-Decoder Reinforcement Learning for Maritime Fleet Renewal under Regime-Switching Market Uncertainty
Maritime fleet renewal requires long-lived investment and contracting under cyclical market uncertainty. This paper develops an integrated artificial intelligence-operations research (AI--OR) framework linking data-driven market uncertainty generation with efficient long-horizon decision optimization and evaluates it through a Very Large Crude Carrier case study. For uncertainty generation, a Gaussian hidden Markov model fitted to historical freight-market index data identifies latent regimes and transitions, while regime-conditional bootstrapping samples market scenarios. Compared with stylized processes or independent sampling, this design preserves market-cycle dynamics and empirical co-movement among freight rates, vessel prices, and market availability limits. The generated scenarios provide the market environment for fleet renewal decisions. For renewal decision making, existing studies often rely on multi-stage stochastic programming solved by rolling-horizon sample average approximation (RH-SAA). However, repeated finite-horizon optimization is computationally costly and may undervalue long-lived actions with short look-ahead windows. We reformulate the problem as a finite-horizon Markov decision process with a strategic-operational decomposition and develop the Post-Decision Optimization-Decoder Reinforcement Learning (POD-RL) framework. POD-RL learns state-dependent latent valuations for post-decision fleet and contract resources, while an optimization decoder converts them into feasible high-dimensional strategic actions. In two 20-year tests under homogeneous and nonhomogeneous state transitions, POD-RL improves mean cumulative profit over RH-SAA by 16.3% and 19.1%, respectively, while reducing online decision time from over 7,000 s to 0.1-0.2 s. When deployed without retraining in a 1999-2018 historical backtest, POD-RL achieves 18.8% higher realized profit than RH-SAA. The results indicate that the integrated framework improves long-horizon decision quality.
- 30 March 2026
- Published by Taylor & Francis in IISE Transactions
- Vol. 59 (2) , 210-233
A hierarchical online planning framework for anticipatory yard allocation optimization in dry bulk terminals
1 IntroductionMaritime transportation is a cornerstone of global trade, with around 80% of international goods transported by sea. Dry bulk cargoes, including coal and iron ore, account for over 40...
- 30 March 2026
- Published by Taylor & Francis in IISE Transactions
- Vol. 59 (2) , 210-233
A hierarchical online planning framework for anticipatory yard allocation optimization in dry bulk terminals
1 IntroductionMaritime transportation is a cornerstone of global trade, with around 80% of international goods transported by sea. Dry bulk cargoes, including coal and iron ore, account for over 40...