One of the main problems in reinforcement learning is to achieve sufficient exploration of the agent's environment. If there is no knowledge about all the reward that is to gather, an optimal decision can hardly be made. However, not leveraging what has been learned so far may slow down the learning progress or make learning unstable.
There exist many exploration strategies to deal with this exploration-exploitation-tradeoff, but they each come with their own strengths and weakensses. The idea of this project is to investigate concepts from model-based reinforcement learning to achieve robust, situation-aware exploration of static and non-static environments.