A CP 2024 Workshop
September 2, 2024
Note: Talks, slides and videos from previous workshops in this series are available here. There is also now a PTHGCP Google Group you can join, for news and discussion related to PTHG topics (not just the Workshop).
Program
Session 1
9:00 – 9:20
Introduction
9:20 – 10:00
Invited Talk: A Constraint Modelling Pipeline
Ian Miguel, Ozgur Akgun
Abstract: This talk will give an overview of the ongoing constraint modelling pipeline project, comprising the Essence language, and the Conjure and Savile Row automated modelling tools, which has grown and developed over the last twenty years. The central idea is to allow a user to specify in Essence a combinatorial optimisation problem without commitment to detailed modelling decisions for a particular modelling paradigm or solver. Alternative models can then be refined automatically from a single Essence specification via Conjure before being tailored to a particular solving paradigm, such as CP, SAT, or SMT, by Savile Row. The talk will also function as a case study of how a long-term, multi-contributor, multi-site project can grow and evolve over time in our efforts to progress towards the holy grail. We will describe fruitful offshoots, such as automated streamlining, constraint-based local search, and instance generation.
10:00 – 10:30
Demo: A Constraint Modelling Pipeline
Ozgur Akgun, Ian Miguel
Abstract: This demo will provide an introduction to the Constraint Modelling Pipeline. We will provide examples of how problems can be specified in the abstract constraint specification language Essence and then refined into a solver-independent constraint model via Conjure before being tailored for a particular solving paradigm and solver by Savile Row. We will demonstrate how Conjure can produce alternative models by following different refinement pathways, and how model enhancements, such as class-wide symmetry breaking, can be added to a model automatically during refinement.
Break: 10:30 – 11:00
Session 2
11:00 – 11:20
Trustworthy and Explainable Decision-Making for Workforce Allocation
Guillaume Povéda, Andreas Strahl, Mark Hall, Ryma Boumazouza, Santiago Quintana-Amate, Nahum Alvarez, Ignace Bleukx, Dimos Tsouros, Hélène Verhaeghe, Tias Guns
11:20 – 11:40
Learning Preferences over Unsatisfiable Subsets
Emilio Gamba, Marco Foschini, Jayanta Mandi, Bart Bogaerts, Tias Guns
11:40 – 12:00
Generalizing Constraint Models in Constraint Acquisition
Dimos Tsouros, Steven Prestwich, Tias Guns
12:00 – 12:30
Panel: Have Chatbots Reached the Holy Grail?
Tias Guns, Thomas Schiex, Christopher Stone
Lunch: 12:30 – 14:00
Session 3
14:00 – 14:40
Invited Talk: Learning how to play (serious) puzzle games, from Sudoku to new functional molecules
Thomas Schiex
Abstract:
Given sufficient data on combinatorial objects generated by an unknown process, such as NP-complete grid puzzles like Sudoku, generative deep learning can approximate the underlying probability distribution and produce new instances. However, this black-box approach is opaque and limited in generating examples with specific properties. We propose a hybrid model combining deep learning and constraint programming. A deep neural network induces a concise rule-based representation of the observed game as a Cost Function Network (CFN). A CFN weighted CSP solver, such as toulbar2, can then generate new instances satisfying additional constraints. This approach effectively learns to play Sudoku and similar games, matching the data efficiency of symbolic methods while offering differentiability for Visual Sudoku learning. We also demonstrate its application to molecular design, outperforming state-of-the-art deep learning-based methods.
14:40 – 15:00
Hedieh Haddad, Pierre Talbot, Pascal Bouvry
15:00 – 15:20
Automating Reformulation of Essence Specifications via Graph Rewriting
Ian Miguel, András Z. Salamon, Christopher Stone
15:20 – 15:30
Community Meeting
Break: 15:30 – 16:00
Session 4
16:00 – 17:30
Lab: A Constraint Modelling Pipeline
Ozgur Akgun, Ian Miguel
Abstract: Following the overview and demo presented earlier in the workshop, in this lab session, participants will gain hands-on experience in using the Constraint Modelling Pipeline. They will be guided through the process of specifying combinatorial optimisation problems using the abstract types provided by Essence and then using Conjure and Savile Row to refine a specification into a model for a target solver. No software download will be necessary, since the lab will be based in Google Colab.
Description
In 1996 the paper “In Pursuit of the Holy Grail” (also here) proposed that Constraint Programming was well-positioned to pursue the Holy Grail of computer science: the user simply states the problem and the computer solves it. It was followed about a decade later by “Holy Grail Redux“, and then about a decade after that by “Progress Towards the Holy Grail“. This series of workshops aims to encourage and disseminate progress towards that goal, in particular regarding work on automating:
- Problem Acquisition: user interaction, learning from examples, model reformulation, debugging, maintenance, etc.
- Solver Construction: tuning parameters, selecting from portfolios, learning heuristics, deep learning, etc.
- User Explanation: reasons for failure, comparison of alternatives, implications for choices, bias detection, suggested modifications, visualization, etc.
Of particular interest is the intersection of the Holy Grail goal with the increasing attention being paid to machine learning, explainable AI, and human-centric AI, and with current work on chatbots and LLMs.
Submissions:
Paper Track:
Submissions may be of any length, and in any format. They may be extended abstracts, position papers, technical papers, or demos. They may review your own previous work or survey a topic area. They may present new research or suggest directions for further progress. They may propose research roadmaps, demonstration domains, or collaborative projects. They may be proposals for measuring progress, and, in particular, for data sets or competitions to stimulate and compare progress. Submit a PDF via an email with subject line: PTHG-24 Paper Track.
Previously Published Track.
Authors are encouraged to submit to this track pointers to relevant papers that they have published elsewhere since the date of the last workshop, PTHG-23, August 27, 2023. The objective is to further the Workshop goal of disseminating progress in this area. Submit a PDF, containing the title, authors, and an abstract of the paper and a link to the paper online, via an email with subject line: PTHG-24 Previously Published Track.
Panel Track:
A panel discussion is planned entitled “Have Chatbots Reached the Holy Grail?”. The answer is presumably “no, but”. Why not? What progress do they embody? How can we get closer? Neuro-symbolic, bespoke chatbots, coping with hallucinations, beyond LLMs, … ? If you would like to participate in the panel, submit a PDF containing a one-paragraph position statement via an email with subject line: PTHG-24 Panel Track.
Submissions should be emailed directly to the Workshop Organizer, at: eugene.freuder@insight-centre.org.
Authors may make multiple submissions if they wish. All submissions that appropriately address the topic of the workshop will be accepted as is, without further revision, and will be made available at the workshop website.
At least one author of every accepted submission must attend the workshop and pay the workshop fees; otherwise the presentation and submission will be withdrawn from the program and proceedings.
Submission deadline: July 4, 2024
Acceptance notifications: July 9, 2024
Organizer:
Eugene Freuder, University College Cork, Ireland, eugene.freuder@insight-centre.org
Gene Freuder 