Tutoring

Automatic Planning (WS 2013/14)

• course website: Automatic Planning

Planning and Learning seminar (WS 2013/14)

• course website: Planning and Learning

I am available for consultations mainly about the "Area 3: Learning policies".

Learning Policies

In automatic planning, a policy is a strategy for solving problems from some planning domain. Reactive policy is usually understood as a function which tells what action to perform in what state. Policies learning is an exciting active research area because having an ideal policy means being able to find the solution to a planning task without the actual need to search.

Specific challenges arise when trying to learn good policies. Although many machine learning algorithms have strong theoretical properties, the actual performance of learning depends greatly on the representativness and coverage of the training data. In practice it may be difficult to generate good training data, as good action choices are usually quite hard to characterize. Consequently, greedy search guided by an imperfect policy often leads to planning failure. The following papers discuss some of these challenges and approaches to solving them.

Reading:

Topic 1: Learning a strategy of action per-domain

• main paper: Approximate Policy Iteration with a Policy Language Bias

Topic 2: Learning weighted rule sets for greedy search

• background: Learning Weighted Rule Sets for Forward Search Planning
• main paper: Iterative Learning of Weighted Rule Sets for Greedy Search