Preference-based Monte Carlo Tree Search

Name: Preference-based Monte Carlo Tree Search

Duration: 11/2016-2019

Funding: DFG

The main objective of this project is the development of preference-based Monte Carlo tree search (PB-MCTS) algorithms, which allow the use of Monte-Carlo tree search in domains where only qualitative feedback is available.

Currently, MCTS methods are limited to numerical rewards, which can not always be assumed. Moreover, numeric feedback might be difficult to extract from natural environments, whereas qualitative feedback may be much easier to obtain because it is considerably easier to determine which of two options is better than to estimate an exact utility value of each alternative