Generalize policy on supporting user scenario

In this thesis we present a way of combining previously learned robot be- havior policies of different users. The main idea is to combine a set of policies, in tabular representation, into a final sub-optimal solution for the problem all users have contributed to. We assume that the features/differences of users are unknown and need to be extracted from the different policies generated from same user. This information is used to weight the importance of a set of actions to sum up two policies. The proposed approach has been tested on a virtual environment finding out that the combined policy works as a general policy suitable for all users, as it always selects actions that are satisfying the users at the border of the defined sensorial possibilities. All the assumptions has been finally verified on a real environment finding out all the limitations of the proposed model.