AI Planning Action Space Reduction via Lifted Successor Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI Planning systems face inefficiencies in handling large domains due to the need to maintain large lists of grounded actions, which is impractical and referred to as 'hard-to-ground', especially when transitioning between different tasks, and combining with Reinforcement Learning techniques is challenging due to data hunger and domain specificity.
Innovation Solution
The implementation of a computer-implemented method that reduces the action space by identifying seed actions through lifted successor generation, treating states as databases and using join queries to generate applicable actions, thereby reducing the size of data structures and improving sample efficiency in Reinforcement Learning policies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a pure Reinforcement Learning system is used, then additional data beyond symbolic model is not required, but the system becomes extremely data hungry and requires adaptation and retraining when moving to different tasks
Solution Approach 1:
The patent combines AI Planning's symbolic model-based approach with Reinforcement Learning techniques. Specifically, it uses the symbolic model to guide the search for solutions while incorporating RL elements, thereby reducing the data hunger of pure RL systems while maintaining adaptability across different tasks through the unified framework
2Device complexity
If AI Planning with symbolic model is used, then additional data beyond symbolic model is not required and it is agnostic to the problem, but it requires maintaining large lists of grounded actions which is impractical for large domains
Solution Approach 1:
The patent segments the action generation process into two parts: (1) using the symbolic model to identify candidate actions without full grounding, and (2) using the learned value function to filter and rank these candidates. This segmentation reduces the size of data structures needed while maintaining the speed of generating applicable actions through the division of computational tasks
3Device complexity
If lifted successor generator is used to avoid grounding, then large lists of grounded actions are not required, but the system converts current state to database and uses join queries which still requires significant data processing
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing value information in the symbolic model before the search process. This preliminary computation allows the system to quickly filter and rank candidate actions during the search without requiring extensive real-time data processing, thereby reducing the processing time for generating applicable actions while maintaining reduced data structure sizes
Data Source
AI summary
Automated improved computer mechanisms are provided for improving the way in which a lifted successor generation (LSG) solution to an artificial intelligence (AI) planning problem is processed. An artificial intelligence (AI) planning problem is received that includes definitions for a plurality of operators. An initial label set, which defines an initial version of an action space, is created, with each label corresponding to an operator. A label reduction is performed on the label set to obtain a reduced label set (seed set) that defines a reduced action space. The AI planning problem is represented as a LSG problem comprising a set of tables and a join query. A LSG module is executed on the LSG problem using the seed set to process the join query and generate applicable action(s) as a solution to the AI planning problem which are then output for further AI operations.