LLM PDDL Generation With EW Feedback for Feasible Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) struggle with generating accurate Planning Domain Definition Language (PDDL) files for planning problems, requiring human expertise and iterative refinement, and small modifications can render plan search infeasible.
Innovation Solution
The use of an Exploration Walk (EW) metric to measure similarity between PDDL domains by comparing executability of random action sequences, enabling iterative generation and refinement of PDDL domain and problem files without human intervention using LLMs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If human expertise and iterative refinement are used to generate accurate PDDL files, then the accuracy and reliability of PDDL generation is improved, but the time consumption and resource intensity increase
Solution Approach 1:
The system enables automated self-refinement where the LLM iteratively generates and improves PDDL files using environment feedback without requiring human intervention. The model automatically evaluates its own output against the environment and refines subsequent generations, transforming a previously human-dependent process into a self-service automated system that maintains high accuracy while reducing time consumption
Solution Approach 2:
The patent implements a feedback mechanism where the system uses environment feedback (execution error messages or returned results) to evaluate and refine PDDL generations. The LLM receives feedback about the executability of generated PDDL files and uses this information to iteratively improve subsequent generations, enabling continuous refinement without human input
2Adaptability or versatility
If small modifications are made to PDDL domain, then the adaptability to different planning problems is improved, but the plan search becomes infeasible
Solution Approach 1:
The system uses environment feedback to detect when modifications to the PDDL domain make plan search infeasible. The feedback mechanism allows the LLM to observe execution results and adjust subsequent domain generations to maintain feasibility while preserving adaptability, creating a self-correcting process that prevents harmful modifications
Solution Approach 2:
The patent implements dynamic refinement where the PDDL domain is iteratively generated and adjusted based on real-time environment feedback. Rather than static domain definitions, the system dynamically adapts the domain structure through multiple generations, allowing flexible modifications while maintaining plan search feasibility through continuous validation
3Manufacturing precision
If iterative generation and refinement of PDDL files is performed, then the accuracy and validity of generated PDDL is improved, but the computational complexity and process duration increase
Solution Approach 1:
The system performs self-evaluation and self-refinement where the LLM automatically assesses the validity of generated PDDL files against the environment and iteratively improves them without external intervention. This self-service approach manages the iterative process internally, reducing the need for complex external coordination while maintaining high precision in PDDL generation
Solution Approach 2:
The patent changes the parameter of generation approach from single-attempt to iterative refinement, where the LLM performs multiple generations and evaluations. This parameter change increases the computational process but is managed through automated parameter adjustment (number of iterations, evaluation criteria) that the system optimizes based on environment feedback
Data Source
AI summary
A method for translating natural language description of planning problems. A natural language description of the planning problem is used to generate problem candidates. Domain candidates are generated from the problem candidates. The domain candidates are refined using an evaluation metric. The problem and domain candidates can be problem planning domain definition language (PDDL) file. A planner can solve the planning problem using the output problem and domain candidates.