Planning AI Assistant for Emergency Decision-Making (PlanAID): Framing Planning Problems and Assessing Plans with Large Language Models
Keywords: Agentic AI, LLMs, Symbolic planner, PDDL
TL;DR: We propose a configurable agentic AI assistant system that utilizes LLMs and a symbolic planner for emergency operations planning.
Abstract: This paper proposes the use of agentic artificial intelligence (AI) to enhance emergency operations planning in response to the escalating frequency and complexity of health emergency response. We present Planning AI Assistant for Emergency Decision-Making (PlanAID), an AI planning assistant that combines a large language model (LLM) with a symbolic planner to improve public health preparedness. Like other re-cent planning tools, ours uses an LLM to translate a planning problem from natural language into a Planning Domain Definition Language (PDDL) specification to be solved by a symbolic planner. We extend this approach by using an LLM with a chat interface to actively help the user 1) frame the relevant components of the planning problem, 2) assess the plan provided by the symbolic planner, and 3) reframe the problem based on issues identified during the plan assessment. This integration allows for the generation of contextually appropriate plans by leveraging the strengths of both LLMs and symbolic planners. Our tool is built on a highly configurable backend that allows an administrator to tailor it for a specific team or incident type by specifying the relevant documents, data sources, plan components, and prompts before an emergency occurs. We also posit evaluation metrics focused on the system's ability to produce effective, resource-optimized plans and its adaptability to complex domains. Our research demonstrates the potential of AI technologies in emergency operations planning, offering a robust solution for improving pub-lic health resilience. Future work will aim to further validate the PDDL outputs and assess the system's practical utility in real-world scenarios, contributing to more effective human-machine teaming and planning processes.
Submission Number: 3
Loading