Main Page

News

Currently, no news are available

Seminar: Trusted AI Planning

Basics. Seminar, 7 graded ECTS points.

The seminar will be run in a block format. There will be an initial meeting on TBA. All student presentations will be given on TBA.

All meetings will take place in room 3.06, Building E1 1. The seminar language is English throughout.

Supervisors for the seminar are Jörg Hoffmann and Daniel Höller. Additional feedback will be provided by TODO.

All email interaction must be pre-fixed with "[TAIP26]" in the email subject.

Your task will be

to read and understand a research topic (see below), to write a summary paper in your own words and to give a presentation.
to provide detailed feedback for the paper and presentation of a fellow student.

No plagiarism, no LLMs. It is Ok (and encouraged!) to use web resources to further your understanding of your assigned topic. However, it is inadmissible to use pieces of public material for your summary paper or presentation, and it is not allowed to generate parts of your slides or summary paper using LLMs. Any plagiarism or LLM use will result in disqualification from the seminar. You are allowed to include pieces (like formal definitions, empirical results tables or figures) from the paper you are summarizing; however, you need to clearly and explicitly mark such material as being from the paper.

Content. AI Planning is the sub-area of AI concerned with complex action-choice problems. The seminar covers methods supporting trust in algorithmic solutions to such problems. This field of research is very recent, and we cover research conducted in the FAI group. A major concern are neural action policies, i.e., neural networks that map states to actions. While such policies can be very performant, they are fundamentally opaque and come without any guarantees. We cover methods for verifying, testing, and re-training such policies.

Prerequisites. Participants must have successfully completed either an edition of the Artificial Intelligence core course, or of the (Trusted) AI Planning specialized course.

Registration. Is via the central seminar registration system.

Grading. The final grading will be based, in this order of importance, on:

The quality of your final presentation.
The quality of your final summary paper.
The quality of the feedback you provide to your mentee student (see below).
Your participation in the discussions during the block seminar.

Summary Paper. For the summary paper, you must use this tex template. You are required to read at least 2 related papers, for the related work section. You are allowed to modify the section structure given in the template if, for whatever reason, this is more adequate for the work you are summarizing.

The seminar paper should be about 4 pages long (not counting the literature list, and in the double-column format of the template). This is a rough guideline, not a strict rule. If you need, say, 5-6 pages to do your topic justice then definitely do so.

Presentation. Your talk should be roughly 20 minutes (not a lot longer, not a lot shorter). It will be accompanied by 10 minutes for discussion. For your presentation slides you are encouraged to use this tex template (or any other tex template).

Schedule and Deadlines (tentative!).

TBA: Initial meeting. We will give a brief insight into each of the papers.
TBA: Choose your preferences (via the Tutorial mechanism in the CMS).
TBA: Receive your topic (as tutorial slots in the CMS). Read the material associated with your topic carefully, and prepare an initial version of your summary paper.
TBA: Deadline for official registration to this seminar (exam registration, Prüfungsanmeldung).
TBA: Meeting with your feedback-giver (as listed with each paper) to discuss your paper. The purpose of this meeting is to ensure that you understood the topic correctly, and to ask questions about specific points.

NOTE: The following deadlines marked with "(ca.)" are meant as a guideline. You are required to do these things, but if you do them 3-4 days earlier or later, that is no problem.

TBA (ca.): Send your summary paper to your mentor student (cc supervisor and feedback giver).
TBA (ca.): Send feedback regarding the summary paper to your mentee student (cc supervisor and feedback giver).
TBA (ca.): Send revised summary paper to your mentor student (cc supervisor and feedback giver).
TBA (ca.): Send presentation slides to mentor student (cc supervisor and feedback giver).
TBA (ca.): Send feedback regarding the revised summary paper to your mentee student (cc supervisor and feedback giver).
TBA (ca.): Send feedback regarding the presentation slides to your mentee student (cc supervisor and feedback giver).
TBA (ca.): Send revised presentation slides to mentor student (cc supervisor and feedback giver).
TBA (ca.): Send feedback regarding the revised presentation slides to your mentee student (cc supervisor and feedback giver).

TBA: Upload your final summary paper in the CMS.
TBA: Upload your final presentation slides in the CMS.
TBA: Block seminar with presentations. Attendance to all talks is required.

Topics. Each participant will be assigned one topic, most of which consists of one paper (while one includes two shortpapers). The overall amount and difficulty of the material associated with each topic is roughly balanced.

Each topic is associated with a mentee student (to whom you will provide feedback, see the deadlines above) and a mentor student (who will provide feedback to you, see deadlines). The mentee/mentor assignment will be a "cycle" through each of the topic areas as listed below: let k be the number of topics, then the mentor->mentee relation is i->i+1 and k->1. If you want to team up with someone specific, please do state that in your email.

NOTE: All papers can be freely accessed from within the university's vpn.

Area 1: Open-Loop Verification (supervisor: Jörg Hoffmannr, feedback giver: Marcel Vinzent)

Topic 1: Guy Katz, Clark W. Barrett, David L. Dill, Kyle Julian, Mykel J. Kochenderfer: Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks. CAV 2017.
Topic 2: Haoze Wu, Aleksandar Zeljic, Guy Katz, Clark W. Barrett: Efficient Neural Network Analysis with Sum-of-Infeasibilities. TACAS 2022.
Topic 3: Gagandeep Singh, Timon Gehr, Markus Püschel, Martin T. Vechev: An abstract domain for certifying neural networks. POPL 2019.
Topic 4: Laurens Devos, Lorenzo Cascioli, Jesse Davis: Robustness Verification of Multi-Class Tree Ensembles. AAAI 2024.

Area 2: Closed-Loop Verification (supervisor: Daniel Höller, feedback giver: Chaahat Jain)

Topic 5: Marcel Vinzent, Marcel Steinmetz, Jörg Hoffmann: Neural Network Action Policy Verification via Predicate Abstraction. ICAPS 2022.
Topic 6: Marcel Vinzent, Siddhant Sharma, Jörg Hoffmann: Neural Policy Safety Verification via Predicate Abstraction: CEGAR. AAAI 2023.
Topic 7: Chaahat Jain, Lorenzo Cascioli, Laurens Devos, Marcel Vinzent, Marcel Steinmetz, Jesse Davis, Jörg Hoffmann: Safety Verification of Tree-Ensemble Policies via Predicate Abstraction. ECAI 2024.
Topic 8: Edoardo Bacci, Mirco Giacobbe, David Parker: Verifying Reinforcement Learning up to Infinity. IJCAI 2021.

Area 3: Testing (supervisor: Daniel Höller, feedback giver: Jan Eisenhut)

Topic 9: Hasan Ferit Eniser, Timo P. Gros, Valentin Wüstholz, Jörg Hoffmann, Maria Christakis: Metamorphic relations via relaxations: an approach to obtain oracles for action-policy testing. ISSTA 2022.
Topic 10: Jan Eisenhut, Álvaro Torralba, Maria Christakis, Jörg Hoffmann: Automatic Metamorphic Test Oracles for Action-Policy Testing. ICAPS 2023.
Topic 11: Maria Christakis, Hasan Ferit Eniser, Jörg Hoffmann, Adish Singla, Valentin Wüstholz: Specifying and Testing k-Safety Properties for Machine-Learning Models. IJCAI 2023.
Topic 12: C. Jain, D. Sherbakov, M. Vinzent, M. Steinmetz J. Davis, and J. Hoffmann: Policy Safety Testing in Non-Deterministic Planning: Fuzzing, Test Oracles, Fault Analysis. ECAI 2025.

Area 4: DSMC, Re-Training, Repair (supervisor: Jörg Hoffmann, feedback giver: Songtuan Lin)

Topic 13: Timo P. Gros, Holger Hermanns, Jörg Hoffmann, Michaela Klauck, Marcel Steinmetz: Analyzing neural network behavior through deep statistical model checking. International Journal on Software Tools for Technology Transfer 2023.
Topic 14: Timo P. Gros, Daniel Höller, Jörg Hoffmann, Michaela Klauck, Hendrik Meerkamp, Verena Wolf: DSMC Evaluation Stages: Fostering Robust and Safe Behavior in Deep Reinforcement Learning. QEST 2021.
Topic 15: J. Eisenhut, D. Fišer, I. Valera, J. Hoffmann: On Picking Good Policies: Leveraging Action-Policy Testing in Policy Training, ICAPS 2025, and
H. Eniser, S. Lin, A. Isychev, N. Müller, V. Wüstholz, I. Valera, J. Hoffmann, and M. Christakis: Using Action-Policy Testing in RL to Reduce the Number of Bugs, SoCS 2025.
Topic 16: L. Cascioli, C., Jain, M. Steinmetz, J. Davis, and J. Hoffmann: Safety Debugging of Tree-Ensemble Action Policies in AI Planning: From Fault Detection to Fault Fixing, RIPL 2025.