Conference Agenda

This work is the fruit of a partnership with Renault. Their inbound supply chain is designed to transport car parts from suppliers spread worldwide to dozens of manufacturing plants through logistics platforms. This logistics operation represents roughly a billion euros and tens of thousands of tons of CO2 emissions per year. Optimizing Renault's logistics network in the long run is therefore crucial. The first step is to optimize the flow of car parts. The resulting problem is a rich version of a multicommodity network flow problem, the load plan design problem. Its computational difficulty stems from its combinatorial nature, the combination of a large time-expanded network, hundreds of thousands of arcs, several millions of commodities to be routed, and the practical necessity to consider bin-packing consolidation explicitly. Although recent resolution methods from the literature scale to very large networks, they do not address more than a fifty thousand commodities and do not consider bin-packing constraints. We model this load plan design problem and propose two tailored heuristics to solve it. The first one bases on a Large Neighborhood Search algorithm, combining existing ideas from the literature and custom perturbations to obtain high-quality solutions. The second one bases on a learned decomposition of the problem to solve it much faster, and responds to an industrial need to interact with the algorithm to test and improve some strategic decisions. We provide a data analysis of Renault's instance and a lower bound to analyze algorithm performance. Our numerical experiments show significant improvement over Renault's current solution.

Timing the introduction of new products to the market is an important strategic decision in the automotive industry. Due to many technical and organization interactions, it is also a complex decision problem that, when approached with mathematical programming, leads to prohibitively long computation times. To adequately support decision makers in advanced what-if analysis, a solution procedure is required that can solve this problem in shorter time and hence allows for its application in a more interactive manner.

In the literature, hybrid algorithms have proven to be adequate solution procedures in such settings, combining features of exploration and exploitation. However, to the best of our knowledge, none of the presented algorithms has addressed the multi-project, multi-mode, multi-criteria resource leveling problem with generalized precedence constraints that forms the structure of the decision problem we study.

In this presentation, we develop a random key genetic algorithm hybridized with a local search routine to solve the problem described above. We provide an overview of its performance when applied to well-known benchmark instances and outline future research directions.

Due to different reasons, the automotive industry is perceiving a massive transition. The shift from ICE to BEV, instable supply chains as well as changed customer demand in the field of digital user experience force car manufacturing companies to adapt their product portfolio more regularly than before. So far, the planning procedure is based on decisions of the involved departments that are partly based on manual calculation of experts, and only partly supported by descriptive analytics. As a conclusion, scenario calculations require long durations and cause high workload.

Hence, we develop a decision support system (DSS) for the strategic product portfolio planning of a car manufacturer based on a MILP integrating the different planning perspectives and requirements. Determining the optimal product portfolio and timing the introduction of the products to the market is one of the most important decisions of the automotive industry. Based on the product architecture, many interdependencies and many interactions between different vehicles must be considered (building blocks, modules, platforms). Moreover, projects rely on shared resources (e.g. financial budgets, development resources, production capacities, etc.), and thus show interdependencies.

The DSS we present allows to calculate and analyze various scenarios based on the formulated MILP. Furthermore, the DSS is designed to be used directly by the decision makers in order to increase data transparency and solution quality. Currently, we are establishing the system within a large car manufacturer in a stepwise process considering different aspects of change management to align to the specific organizational circumstances of the company.