Agent Model Building Program Sensitivity Penalty

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques face difficulties in evaluating conformity to human psychology in agent model building, leading to challenges in improving reproducibility under unknown conditions and measures, as they cannot effectively calculate penalties related to human psychological laws.

Innovation Solution

An agent model building program that selects pairs of samples, acquires output results, and adjusts neural network parameters based on a sensitivity function to ensure training errors and penalties meet predetermined conditions, thereby improving reproducibility.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If a neural network is trained inductively from data without sufficient domain knowledge, then the model can be built without domain knowledge, but the model does not always behave appropriately under unknown circumstances and measures

Engineering Contradiction:
Improveease of model buildingVSAvoidreproducibility under unknown conditions
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent implements feedback by calculating penalties based on sensitivity function violations and using these penalties to adjust neural network parameters during training. The system continuously monitors whether the neural network's output satisfies sensitivity conditions and adjusts parameters to minimize penalty, creating a closed-loop training process that ensures domain knowledge consistency while maintaining inductive learning capabilities

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent changes parameters by adjusting neural network parameters based on calculated penalties from sensitivity function violations. The system modifies network parameters iteratively to minimize the penalty term in the loss function, thereby transforming the network's behavior to conform to domain knowledge requirements while maintaining the flexibility of inductive learning

Inventive Principle:
Principle #35Parameter changes

2Reliability

If a penalty term is added to the loss function to enforce domain knowledge, then the neural network becomes consistent with domain knowledge, but the complexity of designing an appropriate penalty term increases

Engineering Contradiction:
Improveconsistency with domain knowledgeVSAvoidcomplexity of penalty term design
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the penalty calculation into distinct components: selecting sample pairs, calculating neural network outputs, evaluating sensitivity function violations, and computing penalties. This segmentation breaks down the complex penalty term design into manageable steps, where each component has a specific function and can be independently adjusted or replaced based on different domain knowledge requirements

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If existing techniques evaluate physical laws from predicted value structure, then physical law conformity can be checked, but human psychology conformity cannot be evaluated

Engineering Contradiction:
Improveprecision of law conformity evaluationVSAvoidapplicability to different domains
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal sensitivity function framework that can evaluate different types of domain knowledge across various domains. The sensitivity function takes a stimulus variable and response variable as inputs and can be configured to represent different laws (physical, psychological, economic, etc.). This universal structure allows the same penalty calculation mechanism to enforce diverse domain knowledge requirements, making the system adaptable to different application areas

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4191472A1Agent model building program, agent model building method, and information processing device
Publication Date: 2023.06.07 FUJITSU LTD
  • EP4191472A1 patent drawingFigure 1
  • EP4191472A1 patent drawingFigure 2
  • EP4191472A1 patent drawingFigure 3

AI summary

An agent model building program that causes at least one computer to execute a process, the process includes selecting a pair of samples among samples included in a data set, each of the pair of samples having a value other than stimulus variables related to an input to an agent; acquiring a first output result of the data set by inputting the data set to a neural network; acquiring a first penalty based on whether the value conform to a sensitivity function; and adjusting a parameter of the neural network until a training error based on the first output result and the first penalty satisfy a predetermined condition.