Graph Neural Network Agent for Complex System Design
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The design process for complex systems, such as hybrid vehicles, is inefficient due to the time-consuming nature of simulating various configurations to meet functional and non-functional requirements, often requiring extensive manual engineering and separate machine learning models for each system type, which is costly and labor-intensive.
Innovation Solution
A method using a machine learning agent to autonomously create a graph neural network architecture through an iterative process, leveraging reinforcement learning to explore various system designs and learn optimal network structures, allowing for faster and more accurate predictions of key performance indicators by considering unknown dependencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If simulations are used to evaluate system configurations, then functional and non-functional requirements can be verified, but the process becomes time-consuming and computationally expensive
Solution Approach 1:
The patent applies preliminary action by pre-training a neural network model on simulation data before actual design evaluation. The model learns the mapping between system configurations and performance outcomes in advance, so that during the design process, predictions can be made instantly without running full simulations. This pre-computation of evaluation capabilities resolves the contradiction by verifying requirements reliability through the trained model while avoiding time-consuming simulations during the actual design phase.
Solution Approach 2:
The patent creates a copy of the complex system's behavior through a neural network model that replicates simulation outcomes. Instead of running actual simulations repeatedly, the trained network model serves as a lightweight copy that predicts system performance with high accuracy. This copying approach maintains reliability of requirements verification while dramatically reducing the time and computational resources needed compared to running full simulations for each design evaluation.
2Reliability
If manual engineering and separate machine learning models are used for each system type, then specific system requirements can be met, but the process becomes costly and labor-intensive
Solution Approach 1:
The patent implements universality by creating a single multi-functional neural network model that can handle multiple system types and configurations. The model is trained on diverse simulation data covering different system architectures and requirements, enabling it to generalize across various system types. This universal model replaces the need for separate machine learning models for each system type, maintaining system-specific performance reliability while dramatically improving productivity by eliminating repetitive model development and reducing labor costs.
Solution Approach 2:
The patent applies parameter changes by training the neural network model on varied simulation data with different system parameters, configurations, and operating conditions. The model learns to adapt to different system types through parameter variations in the training data rather than requiring separate models. This approach maintains reliability for system-specific performance by capturing parameter dependencies while improving productivity through a single versatile model that can be applied across multiple system types.
3Adaptability or versatility
If the number of configurable options is increased to explore more system possibilities, then better design solutions can be found, but the search space grows exponentially making the process intractable
Solution Approach 1:
The patent introduces an intermediary neural network model that bridges the gap between the vast configuration search space and the simulation evaluation process. The model acts as a mediator by learning the complex relationships between configuration parameters and performance outcomes from simulation data, then using this learned knowledge to guide the search and predict outcomes without exhaustive exploration. This intermediary approach enables comprehensive design option exploration while managing search space complexity by replacing brute-force search with intelligent prediction.
Solution Approach 2:
The patent implements feedback by using the trained neural network model to evaluate design configurations and provide guidance back to the design process. The model's predictions about performance outcomes feed back into the configuration selection process, allowing the system to identify promising designs more efficiently. This feedback mechanism enables thorough exploration of design options by using learned knowledge to prioritize promising configurations while avoiding exhaustive search of the entire exponential search space.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a computer implemented method for providing an agent for creating a graph neural network architecture, which is suitable for providing a prediction of at least one indicator of a complex system and to a computer implemented method for providing such a graph neural network architecture by an agent. Further it relates to an agent and a unit for providing an agent a computer program product and computer readable storage media.