Résumé
The transportation of goods, or city logistics, is a complex system as much by the diversity of its stakeholders (city, transporters, traders, etc.) as by the entanglement of their interactions. Each of these actors has its own objectives, but the effects of decisions to achieve them have repercussions on a global level. Piloting this activity therefore represents a real challenge: not only precisely modeling the system – past or present – poses real difficulties (in terms of complexity and of the data required), but this modeling makes it difficult to predict the effects of the decisions taken on this one. Its management requires the ability to project and represent the effects of decisions, not yet observed, on it. Relevant city logistics decisions must be based on models which must integrate a more detailed vision of activity, at the microscopic level, that is to say operations, towards a macroscopic level which prevailed until then. To do this, we therefore propose to rely on mathematical simulation tools which aim to simulate the macroscopic effects of different decisions (e.g. prohibition of certain engines) by considering microscopic interactions of logistics such as foreseeable change in a logistics organization and the impact on associated indicators (e.g., the level of pollution on a street or the cost of delivery). This is why we are proposing the « open model methodology », the objective is to design and validate a model that can be used by actors in city logistics when making strategic decisions. At the heart of this methodology is the question of integrating expert knowledge into a simulation model. Such a question has agitated the scientific community at least since the creation of AI as a discipline: first with expert systems, whose failure is linked to the impossibility of mechanizing expert knowledge; now with the extremely promising advances in machine learning, which among other things attempt to learn from experts reasoning, but whose models face problems of data availability, validation and explainability. We believe that this methodology makes it possible to reconcile data science and management science so that in complex environments, the decision can be assisted by simulations which allow to master more precisely this complexity. In addition, assuming a certain availability of data, and the desire to set up a data-driven piloting (therefore more automated), this model could serve as a first basis for validating more complex machine learning models.
Source: http://www.theses.fr/2020UPSLM025
.
