The rapid expansion of the on-demand food delivery (OFD) market has led the service providers to manage large-scale dynamic order dispatching. Postponing the dispatch of orders is an effective strategy to alleviate the pressures from sudden order surges and to enhance decision-making quality. This paper addresses the OFD problem with order postponement to minimize travel distances and delayed deliveries, focusing on deciding which orders to postpone and which rider to assign for each remaining order at each decision point. We propose a prior and posterior postponement framework that separates the postponement decision-making process into two phases to balance computational efficiency and decision quality. In the prior phase, multiple knowledge-based postponement rules are designed to quickly filter out orders unsuitable for immediate dispatch. In the posterior phase, a data-driven postponement strategy using reinforcement learning is developed to further optimize long-term objectives. Particularly, an action-oriented phase-specific reward shaping method is designed by analyzing the intrinsic nature of the order postponement process, which helps customize the postponement duration for each order to achieve better postponement performance. Extensive numerical ablation and comparative experiments using real-world data demonstrate that the proposed postponement approach is able to improve customer satisfaction, delivery efficiency, and rider experience better than existing methods. Managerial insights are provided regarding the value of order postponement, key factors for designing effective postponement strategies, and practical ready-to-use postponement tactics.

Pipelining on Edge Tensor Processing Units (TPUs) optimizes the deep neural network (DNN) inference by breaking it down into multiple stages processed concurrently on multiple accelerators. Such DNN inference tasks can be modeled as sporadic non-preemptive gangs with execution times that vary with their parallelism levels. This paper proposes a strict partitioning strategy for deploying DNN inferences in real-time systems. The strategy determines tasks' parallelism levels and assigns tasks to disjoint processor partitions. Configuring the tasks in the same partition with a uniform parallelism level avoids scheduling anomalies and enables schedulability verification using well-understood uniprocessor analyses. Evaluation using real-world Edge TPU benchmarks demonstrated that the proposed method achieves a higher schedulability ratio than state-of-the-art gang scheduling techniques.

This paper considers the cloud-based multi-agent system (MAS) in which for any agent to obtain the most trustworthy data (MTD) that best matches the agents’ personalized demands on the trusted attributes of data, the agents’ data usage often comes with cost, and the better the trustworthiness of data (TOD), the higher the cost of data usage (CDU). Therefore, under the environments where the budgets for data usage are constrained, there may exist a trade-off between the TOD and CDU for the cloud-based MAS, and how to optimally make such trade-off arises as an interesting issue in practice. To address this issue, this paper first formulates the trade-off as a mathematical constrained optimization problem, then explores out the theoretical foundations, both using the fuzzy similarity computing, and finally proposes an algorithm to guarantee the agents to obtain the MTD within the acceptable level of CDU. The experiment results illustrate its effectiveness. Note to Practitioners— This paper addresses the interesting trade-off between TOD and CDU for the cloud-based MAS that operates in the cloud environments where the TOD and CDU are positively correlated with each other. By establishing the mathematical constrained optimization model and designing the algorithm for the MTD recommendation within the acceptable CDU based on the theoretical foundations using fuzzy similarity computing, it helps the MAS with cloud settings to achieve the MTD that best matches their personalized demands on the trusted attributes within the acceptable CDU, which could push the development of cloud-based MAS by benefiting from the optimal trade-off between TOD and CDU. The experiments show that the solution proposed here outperforms existing solutions.