ABSTRACT. AI systems are very useful in practically every sector, but they also have several limitations, mostly related to the lack of reasoning capabilities. According to the fast and slow thinking theory of human decision making, we can say that data-driven AI, including generative AI, are providing fast thinking capabilities, but they do not have slow thinking ones. Existing cognitive theories of human decision making, such as the thinking fast and slow theory, can provide insights on how to advance AI systems towards some of these capabilities.

In this talk I will present a general architecture, called SOFAI, that is based on fast/slow solvers and a metacognitive component that provides a centralized governance of the solvers.

I will describe two instances of this architecture, for constrained grid navigation and planning, showing experimentally that SOFAI generates better decisions than each of the INDIVIDUAL solvers. Emerging behavior related to adaptability, skill learning, and cognitive control are also showed in the analysis of SOFAI’s behavior. I will also describe how the thinking fast and slow theory can help design a value-based human-machine collaborative decision environment.

ABSTRACT. Constraint Programming (CP) is a powerful paradigm for solving complex combinatorial problems, but its adoption is often hindered by the expertise required for modeling. Constraint Acquisition (CA) aims to mitigate this bottleneck by semi-automating the modeling process. This tutorial will provide a comprehensive introduction to CA, covering both passive and interactive learning approaches. For passive acquisition, we will explore CA techniques for learning constraint models from datasets of existing solutions and non-solutions. We will discuss different approaches that focus on learning fixed-arity or global constraints, handling noise, and generalizing the learned models to handle varying problem instances. We will also review interactive CA techniques, highlighting the recent integration of statistical Machine Learning methods that enhance efficiency by reducing the number of queries needed. During the tutorial, state-of-the-art CA tools implemented in the open-source CPMpy modeling language will be demonstrated. Finally, we will discuss current challenges and future directions in constraint acquisition research.

ABSTRACT. This paper explores the application of Maximum Satisfiability (Max-SAT) to the complex problem of conference session scheduling, with a particular focus on minimizing working-group conflicts within the context of the ROADEF conference, the largest French-speaking event aimed at bringing together researchers from various fields such as combinatorial optimization and operational research. A Max-SAT model is introduced then enhanced with new variables, and solved through state-of-the-art solvers. The results of applying our formulation to data from ROADEF demonstrate its ability to effectively compute session schedules, while enabling to reduce the number of conflicts and the maximum number of parallel sessions compared to the handmade solutions proposed by the organizing committees. These findings underscore the potential of Max-SAT as a valuable tool for optimizing conference scheduling processes, offering a systematic and efficient solution that ensures a smoother and more productive experience for attendees and organizers alike.

ABSTRACT. MD5 and SHA-1 are fundamental cryptographic hash functions. Given an arbitrary message, MD5 produces a 128-bit hash in 64 steps, while SHA-1 produces 160 bits in 80 steps. It is computationally infeasible to invert them, i.e. to find a message given a hash. In 2012, 23-step SHA-1 and 28-step MD5 were inverted by SAT solvers, yet no progress has been made since then. In the present paper, it is proposed to construct 31 intermediate inverse problems for any pair of MD5 or SHA-1 steps $(i,i+1)$, such that the first problem is very close to inverting $i$ steps, while the 31st one is almost inverting $i+1$ steps. We constructed SAT encodings of intermediate problems for SHA-1 and MD5 and tuned a SAT solver parameters on the simplest of them. Then the found parameters were used to design a Cube-and-Conquer solver that inverted 24-step SHA-1 and 29-step MD5 for the first time.

ABSTRACT. In the context of aircraft assembly lines, increasing the production rate and decreasing the operating costs are two important, and sometimes contradictory, objectives. In small assembly lines, sharing production resources across workstations is a simple and efficient way to reduce operating costs. Therefore, workers are not assigned to a unique workstation but can walk between them. On the other side, paralleling workstations is an efficient way to increase the production rate. However, the combination of both strategies create complex conditions for tasks to access the production resources. This paper addresses the problem of allocating tasks to workstations and scheduling them in an assembly line where workers can freely walk across workstations, and where workstations can be organized in parallel. We model this novel problem with Constraint Programming. We evaluate it on real world industrial use cases coming from aircraft manufacturers, as well as synthetic use cases adapted from the literature.