Vehicle AI Control with Two-Stage GPT Decision Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle control systems using neural networks and artificial intelligence face challenges such as unpredictability, overfitting, and difficulty in understanding decision-making processes, especially when dealing with diverse vehicle functionalities and goals.
Innovation Solution
A two-stage procedure involving generative pre-trained transformers is employed to operate vehicle functionalities. The first stage generates control data based on a determined goal, and the second stage refines this data to ensure accurate and predictable vehicle operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a general neural network model is used to control diverse vehicle functionalities, then the adaptability and versatility of the system is improved, but the predictability and understandability of the control decisions deteriorates
Solution Approach 1:
The patent divides the control system into multiple specialized neural networks, each responsible for a specific vehicle functionality (e.g., one network for acceleration control, another for braking control). This segmentation allows each network to be simpler and more predictable while collectively handling diverse functionalities, thus resolving the contradiction between versatility and predictability.
Solution Approach 2:
The patent introduces an intermediary layer that translates high-level control goals into specific commands for individual neural networks. This intermediary structure enables the system to maintain predictability at each decision layer while achieving overall versatility through coordinated action of multiple specialized networks.
2Adaptability or versatility
If a large and general artificial intelligence model is used in vehicle control, then the ability to serve diverse goals is improved, but the complexity of the system and difficulty of understanding decisions increases
Solution Approach 1:
The patent segments the artificial intelligence system into multiple smaller, specialized neural networks rather than using one large general model. Each network is trained for a specific vehicle control task, reducing individual complexity while maintaining overall system versatility through the combination of specialized components.
Solution Approach 2:
The patent creates a universal control architecture where multiple specialized neural networks work together under a common framework. This multi-functional system can handle diverse vehicle goals by coordinating the outputs of various specialized networks, achieving versatility without requiring a single complex general-purpose AI.
3Productivity
If machine learning methods are used to control vehicle functions, then the productivity and efficiency of vehicle operation is improved, but the safety and reliability due to unpredictable outputs deteriorates
Solution Approach 1:
The patent divides the machine learning control system into multiple specialized neural networks, each handling a specific vehicle function. This segmentation improves safety by isolating potential failures to individual functions while maintaining overall efficiency through coordinated operation of reliable specialized networks.
Solution Approach 2:
The patent implements feedback mechanisms where the outputs of neural networks are monitored and verified before executing vehicle control actions. This feedback loop allows the system to maintain high productivity through efficient machine learning control while ensuring safety by detecting and correcting unpredictable or erroneous outputs.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In a method for operating a vehicle (102), a goal to be achieved by means of at least one vehicle functionality (FKT1, FKT2, FKT3, FKT4) is determined, and initial input data is generated depending on the goal. A first generative pre-trained transformer (GPT1) is operated based on the initial input data to generate initial output data. Based on the initial output data, a second set of input data is generated. In a first embodiment, the first generative pre-trained transformer (GPT1) is operated based on the second set of input data to generate the second set of output data. Alternatively or additionally, at least a second generative pre-trained transformer (GPT21, GPT22, GPT23) is controlled based on the second set of input data to generate the second set of output data. Finally, the vehicle functionality (FKT1, FKT2, FKT3, FKT4) is operated based on the second set of output data.