-
Background The paper discusses the prospects of integrating Large Language Models (LLMs) into Automated Planning and Scheduling (APS). APS is a branch of AI that involves the development of algorithms and systems to generate plans or sequences of actions to achieve specific goals, while LLMs are neural network models trained on large datasets to understand, generate, and contextualize human language, used in a wide range of applications from text generation to language-driven reasoning tasks.
-
Existing Work Existing APS systems are adept at structured logical planning but often lack flexibility and adaptability to context, a void that LLMs can well fill. Conversely, although LLMs provide unparalleled natural language processing capabilities and a vast knowledge base, they fall short in generating precise, actionable plans where APS systems excel. Integrating LLMs into APS is expected to overcome the limitations of each standalone method, providing a dynamic and context-aware planning approach, also broadening the use of traditional data and past experiences in the planning process.
- Proposed the integration framework of LLMs and APS
-
Challenge 1: Conversion between natural language and structured planning language The literature highlighted the advantages of LLMs in converting natural language into structured planning languages (like PDDL) and vice versa, enhancing the interface between human linguistic input and machine-understandable planning directives.
-
Challenge 2: Lack of capability to generate precise and actionable plans Studies show that while LLMs excel at generating text based on previous input, they face difficulties in producing accurate, detailed spatial reasoning, and processing low-level environmental features. Integrate even imperfect LLM outputs with symbolic planners can speed up heuristic searches, thus enhancing efficiency and reducing search times.
-
The research surveyed 126 relevant papers on the application of LLMs for automated planning, categorized them, and summarized the unique application or customization of LLMs in addressing various planning problems. It was particularly pointed out that in model construction, LLMs have made progress in creating or refining landmarks and domain models. In multi-agent planning, LLMs play a vital role in scenarios involving interaction among multiple agents, enhancing coordination and cooperation for more complex and effective multi-agent strategies. In interactive planning, LLMs are utilized in dynamic scenarios requiring real-time adaptability to user feedback or iterative planning, with outputs refined through different feedback variants.
The paper provides insights into the integration prospects of Large Language Models (LLMs) with Automated Planning and Scheduling (APS), breaking through the traditionally limited adaptability to context, and offers a possibility for a more dynamic, context-aware planning pathway, laying a foundation for further application and research.