Extending Goal-Oriented Specification for Reinforcement Learning to High Dimensions

Deep Reinforcement Learning is used in many complex control problems, such as autonomous driving and control of humanoid robots, but the often complex design of a reward function remains one of its key challenges. Goal-oriented specification addresses this problem and offers an alternative approach to classical reward design. However, it is limited to intuitive and low-dimensional state spaces. This work extends goal-oriented specification by enabling training from high-dimensional state spaces while preserving the intuitive feature space used for goal specification. Using four diverse case studies, we demonstrate that we are not only able to achieve high success rates, but we can also remove unnecessary feature extraction from the deployment of the trained agent. With this, we overcome prior limitations of goal-oriented specification and enable its application to high-dimensional state spaces.

Recommended citation: Extending Goal-Oriented Specification for Reinforcement Learning to High Dimensions; Bjarne Grotelüschen, Simon Schwan and Sabine Glesner; In book: Modeling Decisions for Artificial Intelligence (pp.262-274)Publisher: Springer Nature
Download Paper