Foundations of Artificial Intelligence (FAI) Group
divider line

Seminar: Planning and Learning

Basics. Seminar, 7 graded ECTS points.

The seminar will be run in a block format. There will be an initial meeting on Wednesday, October 16, 10:15--11:45. 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 room 3.06, Building E1 1. The seminar language is English throughout.

The seminar supervisors are Prof. Dr. Joerg Hoffmann, Dr. Peter Kissmann, and Michal Krajnansky.

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 "[PAL13]" in the email subject.

No plagiarism. 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 such material for your summary paper or presentation. Any plagiarism will result in disqualification from the seminar.

Content. Automatic Planning is one of the fundamental sub-areas of Artificial Intelligence, concerned with algorithms that can generate strategies of action for arbitrary autonomous agents in arbitrary environments. A ubiquituous property of planning applications in practice is that the algorithms are run on similar instances -- from the same domain, controlling the same kind of agent in the same kind of environment -- over and over again. Naturally, we want to be able to learn from this experience in order to improve performance over time. The seminar includes a number of recent works in that direction. Specifically, we will cover three different areas where learning is used: performance prediction and portfolio configuration; learning and improving heuristic functions; learning policies (i.e., strategies of action aimed at solving instances from the domain at hand).

Prerequisites. Participants should have successfully completed an introductory course in Artificial Intelligence, and should be familiar with the area of planning to the extent of the material covered in the Artificial Intelligence course we organized this summer (see the lecture slides there).

It is not a necessary prerequisite to have completed the Automatic Planning course, although that is of course an advantage.

Grading. The final grading will be based on:

Summary Paper. For the summary paper, you must use this tex template. Note in particular that you are required to read at least 2 related papers, for the related work section.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.

Topics. Each participant will be assigned one topic, each of which may consists of either a single paper, a part of a single paper, or up to two papers. The overall level of difficulty of the material associated with each topic is roughly balanced. The topics are distributed across the three different areas mentioned above.

Area 1: Performance prediction and portfolio configuration. (Supervisor: Peter Kissmann)

  1. Learning from planner performance: The pioneering investigation on predicting planner performance based on simple features. Paper: Area 1 Topic 1.
  2. Prediction and portfolio configuration: More recent work on prediction, and a follow-up on using prediction for automatic portfolio configuration. Papers: Area 1 Topic 2a; Area 1 Topic 2b.
  3. Portfolio design and analysis: The simple portfolio design mechanisms leading to the winner of the most recent international planning competition; and an alternative design approach and analysis using optimization. Papers: Area 1 Topic 3a; Area 1 Topic 3b.

Area 2: Learning and improving heuristic functions. (Supervisor: Joerg Hoffmann)

  1. Learning heuristic functions by bootstrapping: How to learn distance estimators, with no prior input, by starting with small examples and incrementally going to larger ones. Paper: Area 2 Topic 1 (excluding section 4).
  2. Learning heuristic functions from others: How to learn a better overall estimator when given a set of estimators as input. Papers: Area 2 Topic 2a; Area 2 Topic 2b.
  3. Learning to improve a heuristic: How to learn the difference between the delete-relaxed plan distance estimate, and the real distance. Paper: Area 2 Topic 3.

Area 3: Learning policies. (Supervisors: Michal Krajnansky and Joerg Hoffmann)

  1. Learning a strategy of action per-domain: How to learn a list of decision rules for action selection. Paper: Area 3 Topic 1 (excluding sections 4 and 6 as well as the appendix).
  2. Learning weighted rule sets for greedy search: Related to the previous topic, but in a more refined version prioritizing the rules and using them in a search. Paper: Area 3 Topic 2. As additional background, an earlier simpler version is relevant: Area 3 Topic 2 Background.