Vehicle Decision Support Using Dynamic ML Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing decision-support methods in vehicles, particularly in avionics, do not adequately account for new behaviors that deviate from training data, lacking reliability and accuracy in decision-making support.
Innovation Solution
A method and system utilizing a set of machine learning models with dynamic evaluation and selection, including acquisition, selection, transmission, supply, and return phases, to provide recommendations with associated reliability scores, using prospective, introspective, and retrospective scores for model selection and user feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single machine learning model is used for decision support, then the system is simple to implement, but the reliability and accuracy of decisions deteriorate when new behaviors deviate from training data
Solution Approach 1:
The system segments the decision-making process by maintaining multiple specialized machine learning models (e.g., behavior recognition model, intention prediction model, decision recommendation model) that each handle specific aspects of the decision support task. This segmentation allows the system to achieve higher reliability through model specialization while managing complexity through modular architecture.
Solution Approach 2:
The system merges multiple machine learning models into a unified decision support framework where models work collaboratively. The server combines predictions from multiple models and integrates them with rules from knowledge bases to produce comprehensive decision recommendations, thereby improving reliability through ensemble approaches while presenting a unified interface to users.
2Measurement precision
If multiple machine learning models are applied to evaluate all possible decisions, then the accuracy and reliability of recommendations improve, but the computation time and processing load increase
Solution Approach 1:
The system performs preliminary actions by pre-training multiple machine learning models offline on extensive training datasets. During online operation, the pre-trained models can quickly evaluate decisions without requiring extensive computation time. The server also pre-computes confidence scores and model selection criteria to accelerate real-time decision support.
Solution Approach 2:
The system dynamically selects which machine learning models to apply based on the specific situation, input data characteristics, and required decision type. Rather than rigidly applying all models to all decisions, the system adapts model selection in real-time, applying only the most relevant models to each decision scenario, thereby reducing computation time while maintaining accuracy.
3Productivity
If the system provides automated decision recommendations, then productivity and response speed improve, but the adaptability to new situations and behaviors deteriorates when training data is insufficient
Solution Approach 1:
The system implements feedback mechanisms where decision outcomes are continuously monitored and fed back into the training process. When new behaviors or situations are encountered, the system collects this data, retrains the machine learning models with updated information, and improves future decision-making. This feedback loop enables the system to adapt to new situations while maintaining high productivity through automated operations.
Solution Approach 2:
The system designs machine learning models with universal capabilities that can handle multiple types of decisions and behaviors. The models are trained on diverse datasets covering various scenarios, enabling them to generalize to new situations. The modular architecture allows the same framework to adapt to different vehicle types, operating conditions, and decision domains, enhancing versatility without sacrificing efficiency.
Data Source
Figure 1
Figure 2
AI summary
Method of decision-support for a vehicle or related vehicle simulation, the method being executed by a system comprising a server (20), a client device (10) and a database (30), the method comprising the following phases: a. acquisition phase (100) in which the server (20) receives an input data from a device onboard of the vehicle, b. selection phase (200) in which the server (20) obtain a plurality of machine learning results R1, ..., Rn and computes at least one selection score, then at least one selection score being used to select a preferred machine-learning model MLp, c. transmission phase (300) in which the server (20) sends the result Rp as a recommendation to the client device (10), d. supply phase (400) in which the client device (10) provides the recommendation to a user (60), e. return phase (500) in which the client device returns decision data to a database (30).