Diverse AI Planning Task Generation via Causal Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated planning in artificial intelligence lacks a comprehensive and diverse set of planning tasks, relying on limited hand-crafted domains that do not represent the variety of possible planning tasks, making it difficult to select an effective planner for given problems.
Innovation Solution
The method generates planning tasks with specific causal structures by creating a graph of variables and actions, partitioning atoms into variables, and constructing a layered graph to create a diverse set of planning tasks that can be translated into STRIPS formalism, enabling the generation of a wide range of planning tasks for training and improving planner selection strategies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If hand-crafted domains are used to create planning tasks, then the tasks are easy to understand and control, but the diversity and variety of planning tasks are limited
Solution Approach 1:
The system pre-generates a large library of planning tasks with diverse causal structures before the actual planner selection process. This preliminary generation of varied tasks allows the system to have ready-to-use diverse training data without needing to manually craft each task, resolving the contradiction between task diversity and generation complexity
Solution Approach 2:
The system uses template-based generation where causal graphs are created as reusable templates that can be instantiated multiple times with different parameters. This copying approach enables rapid generation of diverse planning tasks from a small set of template structures, maintaining diversity while reducing the complexity of creating each individual task from scratch
2Reliability
If a large collection of diverse planning tasks is created, then the training data for planner selection is improved, but the time and resources required for task generation increase
Solution Approach 1:
The system performs preliminary generation of diverse planning tasks and stores them in a reusable library. This advance preparation creates a ready-to-use collection of training tasks, eliminating the need to generate tasks repeatedly and reducing the time cost for subsequent training iterations while maintaining high diversity and quality
Solution Approach 2:
The system generates planning tasks by varying parameters such as the number of atoms, variables, and causal graph structures. By systematically changing these parameters, the system can efficiently generate a large number of diverse tasks with controlled variation, improving training data quality without proportionally increasing generation time
Data Source
AI summary
Embodiments for creating planning tasks are provided. A plurality of atoms are generated. The plurality of atoms are partitioned into a plurality of variables. A casual graph is generated based on the plurality of variables. A layered graph including interchanging variable value layers and action layers is created based on the casual graph. A planning task is generated based on the layered graph.