Foundations of Artificial Intelligence (FAI) Group
divider line


Seminar: Neural Networks in AI Planning (NNPLAN)


Basics. Seminar, 7 graded ECTS points.

The seminar will be run in a block format. There will be an initial meeting on Monday, October 25, 16:15-17:45. All student presentations will be given on a single day after the end of term. A detailed schedule is given below.

All meetings will take place in MS Teams . The seminar language is English throughout.

Supervisors for the seminar are Daniel Höller, Daniel Fišer, Marcel Steinmetz, Patrick Ferber, and Jörg Hoffmann

Your task will be to read and understand a piece of research, to write a summary paper in your own words, to give a presentation, and to provide detailed feedback for the paper and presentation of a fellow student.

All email interaction must be pre-fixed with "[NNPLAN21-22]" in the email subject.

No plagiarism. It is Ok (and encouraged!) to use web resources to further your understanding of your assigned topic. Especially, the video presentations of the authors are helpful starting points. However, it is inadmissible to use pieces of such material for your summary paper or presentation. Any plagiarism will result in disqualification from the seminar.


Content. Planning is the sub-area of AI concerned with complex action-choice problems, which occur in a broad range of applications ranging from game playing to smart production. Learning is a natural approach to planning effectively in a given application, and recent results on complex board games (AlphaGo/Zero systems series) has shown the power of this approach. Yet beyond board games this approach is still in its infancy, and strong generalization across structure such as different goals and scaling instance size remains a widely open research problem. The seminar covers works at the current research frontier investigating neural architectures in general planning.


Prerequisites. Participants must have successfully completed an Artificial Intelligence core course. They should be familiar with automatic planning at least to the extent of the material covered in the Artificial Intelligence course; successful participation in one of our AI Planning courses will be an advantage, but is not absolutely necessary to follow the seminar. Furthermore, good basic knowledge on neural networks is required. Ideally, at least one lecture relevant to neural networks was completed.


Registration. Is via the central seminar registration system.


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


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 paper justice then definitely do so.


Schedule and Deadlines (tentative!).


Topics. Each participant will be assigned one topic, each of which consists of one paper. The overall amount and difficulty of the material associated with each topic is roughly balanced.

Area 1: Supervised learning of heuristics (supervisor: Jörg Hoffmann)

  1. Ferber, Helmert, Hoffmann: Neural Network Heuristics for Classical Planning: A Study of Hyperparameter Space. ECAI 2020: 2346-2353.
    Paper: Area 1 Topic 1
  2. Karia, Srivastava: Learning Generalized Relational Heuristic Networks for Model-Agnostic Planning. AAAI 2021.
    Paper: Area 1 Topic 2
Area 2: Graph Neural Networks (supervisor: Patrick Ferber)
  1. Hamilton: Graph Representation Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning. 1-159. Chapter 5. Preprint (2021)
    Paper: Area 2 Topic 1
  2. Shen, Trevizan, Thiébaux: Learning Domain-Independent Planning Heuristics with Hypergraph Networks. ICAPS 2020: 574-584.
    Paper: Area 2 Topic 2

Area 3: Reinforcement learning of heuristic functions (supervisor: Marcel Steinmetz)

  1. Arfaee, Zilles, Holte: Bootstrap Learning of Heuristic Functions. SOCS 2010.
    Paper: Area 3 Topic 1
  2. Ferber, Hoffmann, Helmert: Neural Network Heuristic Functions for Classical Planning: Reinforcement Learning and Comparison to Other Methods. PRL @ ICAPS 2021.
    Paper: Area 3 Topic 2

Area 4: Learning action policies (supervisor: Daniel Höller)

  1. Toyer, Trevizan, Thiébaux, Xie: Action Schema Networks: Generalised Policies with Deep Learning. AAAI 2018: 6294-6301.
    Paper: Area 4 Topic 1
  2. Groshev, Goldstein, Tamar, Srivastava, Abbeel: Learning Generalized Reactive Policies Using Deep Neural Networks. ICAPS 2018: 408-416.
    Paper: Area 4 Topic 2

Area 5: Dynamic Algorithm Configuration (supervisor: Daniel Fišer)

  1. Gomoluch, Alrajeh, Russo, Bucchiarone: Learning Neural Search Policies for Classical Planning. ICAPS 2020: 522-530.
    Paper: Area 5 Topic 1
  2. Speck, Biedenkapp, Hutter, Mattmüller, Lindauer: Learning Heuristic Selection with Dynamic Algorithm Configuration. ICAPS 2021.
    Paper: Area 5 Topic 2