Reinforcement Learning

Reinforcement Learning (RL) is a promising approach to machine learning in which an agent interacts with its environment: The agent sets actions while the environment responds with a change and a numerical feedback. The agent now tries to choose the actions in a way that the cumulative feedback is maximized while going throw the learning task. Since many time-dependent problems can be formulated as a sequence of state, action, and reward, Reinforcement Learning can be used as a solving strategy within this class of problems. The necessary data for the interactions can be provided by a simulation of the model.

Model-Based RL is a special form, where the agent tries to build a functional representation of its environment. This process of model construction can be accelerated by prior knowledge or via special processing steps during interaction. An advantage of this approach is the smaller amount of data needed for successful learning. The scientific challenge is often to approximate the sometimes high-dimensional state-space of the environment appropriately and developing methods to perform this simplification automatically. From a mathematical point of view, convergence guarantees of this learning behavior are very valuable.