Modelling choices in model-based Reinforcement Learning
Speaker: Dr. Georgios Kontes
Affiliation: Self-Learning Systems Group, Precise Positioning and Analytics Department, Fraunhofer Institute for Integrated Circuits IIS
Abstract: Reinforcement Learning (RL) is an area of Machine Learning concerning with agents that take sequential decisions within an environment, aiming at solving a given problem. In the classical Reinforcement Learning paradigm the learning agent has no knowledge of the dynamics of the environment or labeled examples of correct actions, but instead must interact with the environment to design a policy (controller) that maximises future cumulative reward. Among the several different approaches within the RL ecosystem, model-based RL is one of the most sample-efficient, providing also some system safety guarantees. Here, a model of the system dynamics is learned using regression techniques and future actions are selected based on online planning algorithms that utilise the learned model. A core question here is which type of model to use: simpler models might not be expressive enough for complex problems or adequate for large datasets, while higher-capacity models tend to overfit to low-data regimes. Another important aspect is how to consider model uncertainty throughout planning, thus improving data efficiency during learning and safe operation during deployment. In this presentation, different variants from the RL literature for the model and the online planner will be presented.