Note that the STDSM approach is based on decomposition according to the physical structure of the underlying supply chain, i.e., optimization models are solved for the different nodes of the supply chain or groups of them. It is not obvious which one of the two decomposition approaches, i.e. the STDSM or the RBR procedure, is better. It is well-known that planning problems for large-scaled semiconductor supply chains can only be tackled by decomposition (Fordyce et al. 2011). The different planning and control modules are summarized in Table 1. To the best of our knowledge, there is no approach described in the literature that covers the interaction of master planning, allocation planning, and order promising and repromising for semiconductor supply chains. Optimization-based STDSM approaches are not considered so far in the literature for semiconductor supply chains.
Simulation experiments
- Business objects such as orders, lots, machines, and routes are stored in the data layer.
- In both cases, the horizon-based solution obtained from a previous solvewill not be accurate when you move the planning interval.
- Note that the STDSM approach is based on decomposition according to the physical structure of the underlying supply chain, i.e., optimization models are solved for the different nodes of the supply chain or groups of them.
- The experiments demonstrate that the proposed STDSM scheme outperforms conventional business rule-based heuristics with respect to several delivery performance-related measures and with respect to stability.
In this paper, we differentiate between orders that are fulfilled by FP, DREP, and FPOS product aggregates. The structure of the considered semiconductor supply chains including the different product aggregates is shown in Fig. FJSPs.FJSP is a complex class of COPs that involve interdependent assignment and scheduling decisions over extended time horizons, making it more challenging than the basic JSP, which only addresses scheduling Dauzère-Pérès et al., 2023. However, they are typically limited to small-scale instances (fewer than 200 operations) and struggle to scale to real-world, long-horizon scenarios.
3 Simulation results
Computational examples from the fast-moving consumer goods industry are used. An allocation planning procedure for an assemble-to-order (ATO) supply chain is proposed by Chen and Dong (2014). Multiple facilities producing components that are used in various final products are assumed.
The upper value is obtained by the RBR whereas the lower value is computed by the STDSM. Best values for each pair of performance measure values are marked bold. Discussion – RHO for online settings.Unlike offline baselines (CP-SAT, GA, ARD, DRL) that require complete information, RHO constructs solutions progressively using only near-future information, making it suitable for online settings. Next, we explore FJSP variants, including observation noise and machine breakdowns, to evaluate the applicability of L-RHO to batch-online FJSP. The results for the foundry case are similar to the regular cost case. As expected, the OTD and OBD values in Table E-1 of the electronic supplement are slightly better than the ones for the regular case, the same is true for the AWT values since the backlog cost is much higher in the foundry setting.
The approach first segments customers with respect to their importance and profitability into different priority classes. ATP quantities are allocated to these classes based on short-term demand information. Seitz et al. (2020) extend the allocation planning approach of Meyr (2009) by exploiting the known demand forecast bias of customers.
- Finally, the fabrication position (FPOS) aggregate is used to represent the FE level in the supply picture offered by master planning.
- RHO for Long-Horizon FJSP.The temporal structure of FJSP enables the use of RHO to decompose a long-horizon FJSP into a sequence of shorter-horizon FJSP subproblems.
- The semiconductor industry which manufactures integrated circuits (ICs) is one of the most complex industries in today’s world (Mönch et al. 2013).
- The order penetration point is at the interface between planning and control.
- Next, we will describe the different ingredients of the proposed planning approach.
In Table 18, we compare L-RHO with DRL Wang et al., 2023a when training on the same set of 450 FJSP instances. We observe that DRL’s performance degrades with 450 training instances, especially at larger scales, highlighting L-RHO’s unique advantage with a lighter training set. This can be beneficial under situations where FJSP training instances are harder to acquire. Given the above decision variables, the Constraint Programming (CP) formulation for the full FJSP problem P𝑃Pitalic_P is given in Alg. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. ArXiv is committed to these values and only works with partners that adhere to them.
Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling
We now benchmark L-RHO against diverse RHO baselines across FJSP variants. We evaluate objectives such as makespan, start delay, and end delay, while also testing its adaptability under higher system congestion, observation noise, and machine breakdowns to simulate real-world dynamics. A not-for-profit organization, IEEE is the world’s largest technical professional organization dedicated to advancing technology for the benefit of humanity.© Copyright 2025 IEEE – All rights reserved. Use of this web site signifies your agreement to the terms and conditions.
Both the STDSM and the RBR approach require allocation planning, i.e. solving instances of the model (A1)–(A6). Note that the average computing time for a single STDSM decision in the case of the SSC-S supply chain is less than 5 min. Discrete-event simulation is crucial for implementing rolling horizon schemes in a risk-free environment due to the fact that the dynamics and the uncertainty of the supply chain can be covered. Several early papers mention demand fulfillment-related subsystems of semiconductor supply chain planning systems. For instance, a module of the IMPReSS production planning system at Harris Corporation calculates product availability for the quotation and order entry system (Leachman et al. 1996). Requirement and system specification efforts are described by Soares et al. (2000) for an order promising module of a decision support system for semiconductor supply chains, but computational results are not reported.
This is caused by the large number of products, the complexity of the process flows, the difficulty of capacity modeling due to reentrant flows, the size of the production facilities, and the large-sized supply networks in this industry. It is also shown by Mönch et al. (2018b) that demand fulfillment for semiconductor supply chains is an underresearched area. This is at least partially caused by the fact that demand fulfillment strongly interacts with other planning functions which makes it difficult to study it in a stand-alone manner. L-RHO outperforms both traditional solvers (CP-SAT, GA) and the learning-based DRL solver in the standard (offline) FJSP setting.
This makes an introduction to the topic, explains why it is useful, what are some drawbacks and proposes an algorithm to deal with them. Also, it provides two different use cases, lot sizing optimization and tail optimization. This provides an example rolling horizon approach of a crane scheduling problem, with very complete explanations. And also, you could check this very good answer to a question posted some time ago, in which the logic about rolling periods is explained.
Furthermore, when evaluating the performance on 600 operations FJSP (10, 20, 30) in Table 1, we see that option (1) and (2) , results in a longer solve time but an improved makespan from the architecture without attention. We also note that option (3) is strictly dominated by the performance of the architecture without attention. The results of the rolling horizon experiments are shown in Table 5. 95% confidence intervals are presented instead of the values of point estimates to obtain statistically reasonable results. Two values are provided for each factor level and performance measure.
