Abstract
Reinforcement Learning is a highly active research field with promisingadvancements. In the field of autonomous driving, however, often very simplescenarios are being examined. Common approaches use non-interpretable controlcommands as the action space and unstructured reward designs which lackstructure. In this work, we introduce Informed Reinforcement Learning, where astructured rulebook is integrated as a knowledge source. We learn trajectoriesand asses them with a situation-aware reward design, leading to a dynamicreward which allows the agent to learn situations which require controlledtraffic rule exceptions. Our method is applicable to arbitrary RL models. Wesuccessfully demonstrate high completion rates of complex scenarios with recentmodel-based agents.