Title of the paper: Ontology-based Adaptive Reward Functions

Abstract: Reinforcement Learning (RL) is a learning approach where agents receive feedback in the form of a reward function from the environment, allowing them to learn through trial and error. In dynamic environments with unexpected events, there is often a need to design new or adaptive reward functions to dynamically adapt the behavior based on the changing dynamics of the environment. Current methods for specifying reward functions are limited to manual reward function definition or extracting/inferencing from human demonstrations. On the other hand, ontologies with the ability to provide a structured representation and organize concepts and properties hierarchically, facilitate a deeper understanding of the environment, empowering agents to identify and comprehend new events. This paper presents a new Ontology-based Adaptive Reward Function (OARF) method, which dynamically creates new reward functions based on domain ontologies. The OARF method is evaluated in a job shop scheduling environment, demonstrating its superiority over a state-of-the-art baseline algorithm. The evaluation shows improved resource utilization rate, total processed orders, decreased average waiting time, and total failed orders, highlighting the effectiveness of the OARF method.