Goal directed design of rewards and training features for self-learning agents in a human-autonomy-teaming environment

Abstract

For architects of reinforcement learning (RL) agents in real-world applications, the design of a suitable training environment is challenging. Feature engineering and reward function design are mainly guided by human intuition and domain knowledge. We propose the application of a goal directed task analysis (GDTA) to structure knowledge about an application domain in order to guide the design of a RL agent embedded in a complex environment. Results from the task analysis can be leveraged for feature selection and reward function design. We showcase this approach in a human-autonomy-teaming application for military airborne operations.