Vortrag: Introduction to Reinforcement Learning
Reinforcement Learning is a powerful approach to machine learning which is based on experience without prior knowledge or guidance from experts. It enables an AI to independently create models of its environment and develop appropriate action strategies for goal-oriented tasks. The self-learning algorithms can be applied to time-dependent problems in a changeable and unknown environment. Applications include Game Ai (Alpha Go / Alpha Go Zero), Real-time Decisions and Robot Navigation.
The aim of the lecture is to provide an insight into the theoretical and conceptual fundamentals of Reinforcement Learning, as well as a basic understanding of the best-known RL algorithms (SARSA, Q-Learning,..).
The formal framework of the "Markov Decision Processes" will be discussed, allowing time-dependent and decision-based tasks to be represented in a meaningful way: The learning task is modelled as an interaction between the "environment" and an "agent" acting in it. The goal of the agent is to find strategies to maximize a previously defined reward through the environment. Thereby goal-oriented behavior can be translated as optimization problem, which is approximately solved by the experience gained by the agent.
The methods are demonstrated with simple applications from deterministic and stochastic environments with code examples.