The project aims to develop an open domain dialog management system using knowledge-based graphs. These graphs are principled data structure for knowledge representation that can be easily extended by new kinds of information, therefore it supports rapid deployment of new domains in dialog systems which use the knowledge-based graphs. The extension of current systems is difficult since the used semantic and knowledge representation is typically restricted to a specific domain. The second goal is to develop algorithms for dialog management optimization by using machine learning from data collected directly during the dialog in order to simplify deployment of the systems in real-life scenarios. Machine learning is already used for purposes of dialog management, however system is usually trained without direct interaction with the user. The benefit of our project is in creating easily extensible dialog system, which deployment will be simplified by using algorithms for getting training data interactively from the users. The third goal is to integrate the developed algorithms with a dialog system which is used by real users.