A Survey of Non-Learning-Based Abstractions for Sequential Decision-Making

Abstraction techniques play a crucial role in enabling agents to make decisions more effectively by simplifying complex problems. This survey provides a comprehensive literature overview of non-learned abstraction construction methods and explores how these abstractions can enhance or be seamlessly integrated into existing solvers. We delve into key properties of abstractions, outline general strategies for constructing them, and discuss specialized approaches for specific problem domains, such as those with continuous action spaces. Additionally, we introduce the Abstraction Mapping Graph (AMG) framework, offering a structured lens through which abstraction usage can be systematically understood.