Safe ML Model Training Using Multitask Gaussian Processes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing safe learning algorithms for machine learning models in engineered processes and machines face challenges in accurately modeling safety values and tend towards local exploration, requiring prior knowledge and domain expert input.
Innovation Solution
The method employs a multitask Gaussian process to implement a joint model of safety values for both the target and auxiliary systems, allowing for the prediction of safety values and enabling global exploration across disjoint safe regions using transferable auxiliary knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If safe learning algorithms use Gaussian processes to model safety values and limit exploration to the current safe set, then safety confidence is maintained, but the exploration is restricted to local regions and disconnected safe regions cannot be discovered
Solution Approach 1:
The method pre-computes and stores auxiliary data from disconnected safe regions before the main learning process begins. This preliminary action allows the system to later jump to these pre-identified safe regions without needing to explore through unsafe areas, thus maintaining safety confidence while expanding exploration coverage to include disconnected regions that would otherwise be inaccessible.
2Adaptability or versatility
If domain experts supply safe data from multiple safe regions, then the safe learning algorithm can explore more regions, but the process becomes more labor-intensive
Solution Approach 1:
The system automatically identifies and explores disconnected safe regions by using the pre-computed auxiliary data and the multitask Gaussian process model, eliminating the need for domain experts to manually supply safe data from multiple regions. The algorithm performs self-directed exploration by leveraging the auxiliary data to guide it to disconnected safe regions, thereby reducing labor intensity while maintaining comprehensive exploration coverage.
3Productivity
If a multitask Gaussian process jointly models safety values for target and auxiliary systems, then data consumption is reduced and learning is accelerated, but computational load increases
Solution Approach 1:
The method pre-computes and stores auxiliary data from the auxiliary system before the main learning process. This preliminary computation allows the multitask Gaussian process to leverage pre-processed information during training, reducing the computational burden of jointly modeling both systems in real-time while still achieving accelerated learning and reduced data consumption through the transfer of auxiliary knowledge.
Solution Approach 2:
The method extracts and separates the auxiliary data computation from the main target system training. By computing auxiliary data independently in advance and then integrating it through the multitask Gaussian process, the system reduces the real-time computational load while maintaining the benefits of joint modeling for accelerated learning and reduced data requirements.
4Measurement precision
If safe learning algorithms require precise safety models calibrated before exploration, then safety constraints are accurately modeled, but the setup process becomes more complex and time-consuming
Solution Approach 1:
The method performs preliminary computation of auxiliary safety data from the auxiliary system before the main exploration begins. This pre-computed auxiliary data serves as a foundation that reduces the need for extensive calibration of the safety model for the target system, thereby maintaining accurate safety constraint modeling while significantly reducing the setup time and calibration effort required.
Data Source
AI summary
A method for training a target machine learning model for a target system in engineered processes and machines. A multitask Gaussian process implements a joint model of safety values of the target and auxiliary system. A new state is selected for the target system, wherein target safety values are predicted by the multitask Gaussian process.