Multitask Gaussian Process Training for Safe Global ML Exploration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing safe learning algorithms for machine learning models face challenges in accurately modeling safety values, especially in scenarios where prior knowledge is limited, leading to local exploration and increased data consumption.
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 of disjoint safe regions with reduced data consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If safe learning algorithms use Gaussian processes to model safety values with prior knowledge, then safety confidence is improved, but data consumption increases
Solution Approach 1:
The patent applies preliminary action by pre-training a Gaussian process model on auxiliary system data before actual safe learning begins. This pre-computation of safety models from auxiliary observations enables the target system to start with prior knowledge, reducing the need for extensive data collection while maintaining safety confidence. The auxiliary training data is processed in advance to establish initial safety boundaries.
Solution Approach 2:
The patent uses an auxiliary system as an intermediary to transfer safety knowledge to the target system. The auxiliary system serves as a mediator that provides pre-collected safety observations, which are then used to initialize the Gaussian process model for the target system. This intermediary approach reduces direct data consumption from the target system while maintaining reliability through transferred knowledge.
2Reliability
If safe learning algorithms explore only regions with high safety confidence, then safety is improved, but exploration scope is limited to local regions
Solution Approach 1:
The patent applies dimensionality change by transferring knowledge from an auxiliary system dimension to the target system dimension. Instead of exploring only locally in the target system's state space, the algorithm leverages observations from the auxiliary system (a different dimension) to expand exploration scope. This cross-dimensional knowledge transfer enables global exploration while maintaining safety guarantees.
Solution Approach 2:
The patent uses preliminary action by pre-establishing safety models from auxiliary system observations before target system exploration begins. This advance preparation creates a broader safety confidence map that enables exploration beyond immediate local regions, as the pre-computed model provides prior knowledge about safe regions that would otherwise require extensive exploration to discover.
3Measurement precision
If precise safety models are required before exploration, then safety accuracy is improved, but model calibration complexity increases
Solution Approach 1:
The patent uses the auxiliary system as an intermediary to provide pre-calibrated safety observations, eliminating the need for complex model calibration in the target system. The auxiliary system serves as a proxy that has already performed the calibration work, and its observations are directly transferred to initialize the target system's safety model, reducing both complexity and time requirements.
Solution Approach 2:
The patent applies copying by replicating the safety model structure and initializing it with observations from the auxiliary system. Instead of calibrating a new model from scratch for the target system, the approach copies the model framework and populates it with transferred knowledge, significantly reducing calibration complexity while maintaining safety accuracy through the copied structural constraints.
4Reliability
If Gaussian processes are used for safe learning, then safety confidence is improved, but computational load increases
Solution Approach 1:
The patent applies preliminary action by pre-computing the Gaussian process model on auxiliary system data before target system operation. This advance computation transfers the heavy computational burden to an offline phase, allowing the online target system to use the pre-computed model with minimal real-time computational load, thus maintaining safety confidence while reducing energy consumption during operation.
Solution Approach 2:
The auxiliary system acts as an intermediary that absorbs the computational load of Gaussian process training. By performing the computationally intensive model calibration on auxiliary data rather than target system data, the approach transfers computational burden to a separate system, reducing the energy requirements for the target system while maintaining safety confidence through the transferred model.
Data Source
Figure 1a~1b
Figure 2a~2b
Figure 3a~3c
AI summary
Some embodiments are directed to 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.