Using Machine Learning Techniques in Search

A typical ability used by people in problem solving is to learn from prior experiences in problem solving and to use the learned knowledge to produce better results or solve harder problems than they would be able to achieve without this learned knowledge. Obviously, especially knowledge-based search should profit very much from the use of machine learning techniques, both to select good suited search approaches and control strategies for the given problem and to guide a system during its search. But there is one important problem that has to be taken into account when developing learning search systems: usually the search approaches employed by computer systems are not the same problem solving approaches that people use and therefore it is not obvious how the learning process of people can be transfered to the search approaches.

When trying to develop a learning search system, we have found the following questions to be a good guideline for the development. The first four questions are concerned with the learning phase of the system:

• What or whom do we learn from?
• What to learn?
• How to represent and store the learned knowledge?
• What learning method to use?

But learned knowledge can only help a search system if the system can apply this knowledge during its search. This is therefore called the application phase of the system and there are also four questions to guide the development of this part of it:

• How to detect applicable knowledge?
• How to apply knowledge?
• How to detect and deal with (applicable, but) misleading knowledge?
• How to combine knowledge from different sources?

Finally, at the bottom of all the questions presented so far, there is an elementary question that all learning systems (not only search systems) have to answer:

• Which concepts of similarity are helpful?

We have developed various concepts for learning search systems that are based on different answers to several of the questions from above. There are own web-pages for the following approaches:

• learning the selection of a good control
• flexible re-enactment of successful search steps for similar problems
• learning abstracted pattern of successful search steps

An overview of all the known answers to the nine questions for the application area automated deduction is given in the following report:

• Denzinger, J. ; Fuchs, M. ; Goller, C. ; Schulz, S.:
  Learning from Previous Proof Experience: A Survey,
  AR-Report AR-99-4, TU München, 1999.
  • Abstract
  • Download full paper (187 Kbytes)

to our page on learning the selection of a good control.

Last Change: 5/12/2013