TB8* - Network Design-Data Visualisation

Cette présentation explore la manière dont les données générées au cours d’un processus d’écriture peuvent être collectées, analysées et interprétées pour mieux comprendre les dynamiques de création textuelle. Grâce à une combination des fichiers d’enregistrement de frappes et la théorie des graphes, nous mettons en lumière les épisodes de modifications du texte en construction.

Mots clés : processus d’écriture, graphe, communauté, données d’écriture.

Writing is a complex task that involves a set of basic opérations through which improvements are achieved to reach the desired text quality. Graph models have been proposed since 2011 to visualize this process. With the use of LLMs, the approach to build a text changes. Instead of writing for a potential reader, the writer then writes for the computer to generate a text designed for the final potential reader. Various strategies may be used. Some avenues to study this new form of writing are explored or proposed.

Building on the consolidation-based framework of Hewitt and Lehuédé, we present a column generation approach for CTSND that embeds departure times into consolidation variables, reducing the pricing problem from quadratic to standard knapsacks while strengthening the LP relaxation. Computational results on benchmark instances demonstrate stronger lower bounds and competitive runtime.

Large-scale deterministic network design problems remain computationally challenging, as exact Benders-based methods often struggle to scale. This paper proposes a demand-sampling framework for problems whose objective contains a large sum of demand-specific routing terms. The key idea is to sample demands and reweight their contribution so that, for any fixed design, the sampled objective remains unbiased for the full deterministic objective. This makes it possible to solve sampled approximations within a Sample Average Approximation framework and to obtain, in expectation, a lower bound on the true optimum. We develop two sampling schemes with theoretical guarantees and pair the resulting solutions with statistical bounds on solution quality. We show the effectiveness of our approach on large benchmark instances of the Traveling Salesman Problem with Generalized Latency and the Multicommodity Uncapacitated Fixed-Charge Network Design problem.