categorisation

Categorisation

This micropsi sub-project is mainly about empowering the agent to learn two different kinds of hierarchies within its representation of the world, namely partonomies (part-of hierarchy) and taxonomies (is-a hierarchy).

Methods being tested for learning a partonomy are categorical Boltzmann machines for means of statistical learning and learning by interaction with the object in the world.

Methods being tested for learning a taxonomy are more diverse, including knowledge based artificial neural networks for supervised learning and an clustering approach based on a similarity measure for object representations in node nets for unsupervised learning. Combinging supervised and unsupervised learning methods this framework should in the end allow a semi-supervised learning of taxonomies.

Current Activities

Right now, we're restructuring the code for better readability and performing the first longer test-runs of our partonomy learning system. We hope to find some sensibvle parameter values soon. However, test-runs do still not last long enough due to problems sending the agents actions to the world.

People involved

• mischu
  mainly concerned about learning of partonomies
• colin
  works on his diploma about learning of taxonomies using the micropsi toolkit.
• gregor
  

External Links

• http://www.informatik.hu-berlin.de/~mhimmels/microPsi/
  Page devoted to categorisation and learning within the micropsi toolkit.