Oct 2026· Tsinghua Science and Technology· 0 citations
TL;DR
A deep reinforcement learning–based algorithm is developed at the high level, which employs an attention-based model to effectively exploit both the current environment state and historical visiting information when selecting the next tar-get.
Abstract
This paper investigates an uncertain Moving-Target Traveling Salesman Problem (MT-TSP), in which the future motion trajectories of targets are uncertain. The objective is to maximize the number of visited targets within a bounded region while minimizing the total travel distance. To address the challenges introduced by dynamic target maneuvers, we propose a two-level online solution framework, termed PaE (Plan and Execute), which leverages the current observations of targets to solve the problem in real time. The high-level module plans the next target to visit, while the low-level module executes the corresponding access process. We develop a deep reinforcement learning–based algorithm at the high level, which employs an attention-based model to effectively exploit both the current environment state and historical visiting information when selecting the next tar-get. At the low level, rapid target access is achieved through a lightweight prediction mechanism with continuous state updates. Extensive experiments in simulated environments demonstrate that PaE is capable of extracting key problem features to capture dynamic changes, leveraging rich contextual information to guide decision-making, and generating high-quality solutions within extremely short computation times.
Few-shot coverage and capacity optimization (CCO) under a small adjustment budget is challenging: model-based methods suffer from simulator mismatch, while feedback-based methods are sample-starved. We address this challenge with a rollout-based policy-improvement framework. Rather than deploying the trained policy's h...
Jin-Dan Li, Samuel Albert, Yan Xin et al.· ACM Interational Symposium o...· 0 citations
This work proposes KnowGPO, a reinforcement learning framework that aligns trajectories by the task-relevant knowledge they have acquired rather than by their surface observations, and demonstrates how this helps solve knowledge-intensive tasks through long-horizon, multi-turn interactions.
Si-Qi Chen, Zhengxi Lu, Yuchen Yan et al.· Proceedings of the 2026 Conf...· 0 citations
In this work, we consider decision-focused learning (DFL) for a Markov decision process (MDP), where existing methods differentiate through the KKT conditions of the Bellman equation and require solving a linear system over all state-action pairs, limiting its scalability. We address this by reformulating the MDP as an...
Zi-Hao Zhao, Ashwath K. Karunakaram, A. Eshragh et al.· 0 citations
This work introduces Trajectory-to-Step Policy Optimization (T2SPO), a method that uses past interaction trajectories to provide step-level feedback for policy learning and shows that T2SPO consistently improves overall task success over GRPO.
Bo-Wen Zhang, Jun-Wei He, Mao-Qi Liu et al.· 1 citation
The proposed hierarchical multi-agent proximal policy optimization framework can reduce total airlines' operational costs—including direct operating cost and capital cost and achieves a computation speedup in comparison with a conventional optimization baseline.
Li-Jing Liu, James M. Shihua, Qi-Yu Yan et al.· MATEC Web of Conferences· 0 citations
LP-BTS is a learning-guided planning architecture that exceeds a deadline heuristic and two source-derived direct-policy reconstructions on every paired scenario, and provides controlled evidence about learning-guided planning in a large, dynamic action space.
Liang-Ching Tao, Pi-Chung Wang· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.