Research

I am interested in inductive learning in automatic planning and machine learning in general. Currently I am working on the topic "Learning Bad Action Choices in Automatic Planning".

Learning Bad Action Choices

There are many approaches to applying machine learning techniques in automatic planning. One of the most studied approaches is learning domain control knowledge in the form of reactive policies, which define good action choices. It turns out that good action choices are quite hard to characterize in the domain-independent manner. Induced knowledge is always as good as the training data, and the learned policies are almost never perfect. This consequently leads to planning failure when using the learned policies greedily, and robust integration of the learned policies in search itself poses a challenge.

On the other hand, at least some subset of bad action choices can be characterized quite easily. Also, the learned knowledge can be integrated quite straightforwardly in some of the existing planning systems as pruning rules.

As classical planning deals with the logical description of the planning domains and tasks, the inductive logic programming (ILP) approach to learning was chosen as a way-to-go. The rules generated by the inductive logic learner are easily interpretable and verifiable by human. Also, the expresivity of the hypothesis language is only constrained by the expresivity of the first-order predicate logic and the practical limitations on the learning and evaluation efficiency.