Deep Learning Component Selection for Satisficing Planners

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional satisficing planners lack an efficient method to determine the optimal combination and order of components, such as heuristic functions, search algorithms, and search refinement algorithms, leading to suboptimal performance in various planning tasks.

Innovation Solution

Utilize a machine learning algorithm, specifically deep learning, to design and select the appropriate components and their order for a satisficing planner based on application contexts, leveraging trained models to determine the most suitable search algorithm and search refinement components.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional satisficing planners use manual component selection, then the planner structure is simple and easy to understand, but the performance is suboptimal due to inability to determine optimal component combinations

Engineering Contradiction:

Improveplanner performanceVSAvoidcomponent selection complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The system uses deep learning models to automatically select and configure planner components based on task characteristics, eliminating the need for manual expert configuration. The model self-learns optimal component combinations from training data, enabling autonomous decision-making for component selection.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The approach transforms the discrete component selection problem into a continuous parameter optimization problem by using neural network outputs to determine component configurations. The system learns optimal parameters for component selection through training on planning tasks.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If deep learning is used to optimize component selection, then planner performance improves, but the system complexity increases due to additional training and inference components

Engineering Contradiction:

Improveplanning efficiencyVSAvoidsystem architecture complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The deep learning model is trained in advance on a dataset of planning tasks and their optimal solutions. This preliminary training phase enables the model to learn patterns and make rapid component selection decisions during actual planning execution, separating the complex learning process from the operational phase.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The deep learning model acts as an intermediary between the planning task description and the component selection process. It translates task characteristics into optimal component configurations, decoupling the complexity of component selection from the planning algorithm itself.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12412071B2Creating satisficing planners with deep learning

Publication Date: 2025.09.09 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

Some embodiments of the present invention are directed to a method of choosing the components for a satisficing planner using machine learning (ML) (for example, deep learning (DL)). Some embodiments of the present invention are directed to choosing search algorithm component(s) for a satisficing planner using ML (for example, DL). Some embodiments of the present invention are directed to choosing search refinement components (that is, search boosting component(s) and/or search pruning component(s)) for a satisficing planner using ML (for example, DL).