PhD position in Explainable Combinatorial Optimisation (TU Delft, The Netherlands)

[Emir Demirović]

We are seeking a highly motivated candidate to work on the intersection between combinatorial optimisation, AI, and explainability. The goal of the project is to develop novel combinatorial optimisation methods that not only exhibit excellent performance, but also concisely break down the main factors behind automated decision-making in a way that is easy for humans to interpret.

The ideal candidate has excellent algorithmic and programming skills, as demonstrated through projects or programming competitions. Knowledge on formal reasoning, logic, and SAT solving is highly desirable, as well as knowledge on operations research and constraint programming. Candidate lacking skills in one area may make up for it with a strong background in another area. Willingness to contribute to teaching is also important.

The candidate will join the newly formed XAIT Delft AI Lab alongside three other PhD students. The aim of the lab is to develop novel approaches that focus on the explainable aspect of artificial intelligence for transportation and other civil engineering problems and promote AI-related education. The position is for five years. The extra year compared to the usual four-year contract accommodates the 20% additional AI-related education related activities.

The PhD student will be supervised by Dr Emir Demirović and Dr. Gonçalo Correia at TU Delft (The Netherlands).

The deadline for applications is August 21. Interviews will follow shortly after.

For more details, please see the formal ad: https://www.tudelft.nl/over-tu-delft/werken-bij-tu-delft/vacatures/details?jobId=7687&jobTitle=PhD%20position%20in%20Explainable%20Combinatorial%20Optimisation

[Gupta, Gopal]

Dear Colleagues,

Since I see a lot of traffic here about explainability in AI, I thought I will mention these 2 tools for explainable AI that my group has developed. These tools will learn a symbolic model expressed as a default theory---a stratified answer set program (and ASP is closely related to constraint programming). These are industrial strength tools that produce an interpretable model in which a prediction can be explained. They are competitive with state-of-the-art traditional machine learning tools such as XGBoost or Multilayer Perceptrons (MLPs), however, in addition, our systems generate models that are interpretable/explainable. They are freely available on Github. They have been adopted by Atos, the French Software giant, as part of their XAI toolchain. The great thing about these tools is that complex data transformations such as one-hot-encoding are not needed. High school students with a little knowledge of python have used them successfully. These tools are also super-efficient (just takes a few seconds to output the model).

FOLD-R++ does binary classification. FOLD-RM does multi-category classification.

The papers describing the tools and their performance are here: FOLD-R++: https://arxiv.org/abs/2110.07843 (appears in FLOPS'22)

                                                                                                                    FOLD-RM: https://arxiv.org/abs/2202.06913 (appears in ICLP'22)

 

The tools are freely available on github: FOLD-R++: https://github.com/hwd404/FOLD-R-PP

                                                                        FOLD-RM: https://github.com/hwd404/FOLD-RM

Enhanced version of these tools will be available soon. The enhanced versions produce even smaller number of rules, produce more or less the same rules regardless of how the data is split between testing and training, and are more efficient.

As an example, for the well-known Titanic dataset on Kaggle, just 2 rules generated by FOLD-R++ suffice to produce 98.6% accuracy.

      status(X,perished) :- not sex(X,'female').
      status(X,perished) :- class(X,'3'), sex(X,'female'), fare(X,N1), not(N1=<23.25).

and obviously

      status(X,survived) :- not status(X,perished).

If you achieve something of significance with these tools, please write to me.

Limitation of the tools: only work with tabular data containing numerical and categorical attributes (no images).

Enjoy

Gopal Gupta

https://utdallas.edu/~gupta

---

From: const...@googlegroups.com <const...@googlegroups.com> on behalf of Emir Demirović <emir.de...@gmail.com>
Sent: Friday, July 15, 2022 12:13 PM
To: const...@googlegroups.com <const...@googlegroups.com>
Subject: [constraints] PhD position in Explainable Combinatorial Optimisation (TU Delft, The Netherlands)

 

--
You received this message because you are subscribed to the Google Groups "Constraints" group.
To unsubscribe from this group and stop receiving emails from it, send an email to constraints...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/constraints/CAM7BFG8NTX2LkeXRtSjQctacE0gFWDHZ2L7r7gq8r3iQWJ%3D17w%40mail.gmail.com.