AutoML Pipeline Generation with Dynamic Probability Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AutoML technologies lack the ability for users to choose custom generation strategies for machine learning algorithms (MLAs) and often employ non-optimized strategies with limited interpretability and transparency, making them unsuitable for generating ML pipelines for large-scale data center operations within limited time and resource constraints.
Innovation Solution
A computer-implemented method for automatically generating a target machine learning pipeline by accessing candidate ML pipelines and variation operators, iteratively updating the pipelines based on training data, and dynamically adjusting the probability distribution functions of the variation operators to select the most effective operators, resulting in an optimized MLA ready for production.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If typical AutoML technologies are used to generate machine learning algorithms, then the generation process is automated, but the generation strategy is non-optimized and lacks interpretability
Solution Approach 1:
The patent implements feedback mechanisms where performance scores from evaluating candidate MLAs are used to dynamically adjust probability distribution functions of variation operators. This feedback loop enables the system to learn from previous iterations and optimize the generation strategy automatically, resolving the contradiction between automation and optimization by making the strategy adaptive rather than static.
Solution Approach 2:
The patent introduces dynamic probability distribution functions that change over time based on performance feedback. Instead of using fixed uniform distributions, the system dynamically adjusts the likelihood of selecting specific variation operators based on their past performance, transforming the static generation process into a dynamic, self-optimizing system that maintains automation while improving effectiveness.
2Ease of manufacture
If typical AutoML technologies iteratively select variation operators according to uniform distributions, then the generation process is simple, but the strategy is not optimized for performance
Solution Approach 1:
The patent changes the parameters of the selection process by transitioning from uniform probability distributions to dynamically adjusted probability distributions. This parameter change allows the system to favor variation operators that have demonstrated better performance while maintaining the iterative generation process, thus improving productivity without fundamentally altering the overall generation methodology.
Solution Approach 2:
The system performs self-service by automatically adjusting its own generation strategy based on performance feedback. The probability distribution functions are dynamically modified using performance scores from previous iterations, enabling the system to self-optimize without external intervention, thereby improving effectiveness while maintaining simplicity.
3Measurement precision
If more data is collected for training ML models in data centers, then model accuracy improves, but the time and processing power required increases
Solution Approach 1:
The patent applies partial action by using performance scores from a subset of evaluation data to guide the optimization process. Instead of requiring full training on all available data for every iteration, the system uses efficient performance metrics to guide the search, reducing the time and computational resources needed while still achieving accurate models through the optimized generation strategy.
Solution Approach 2:
The system performs preliminary action by pre-evaluating candidate MLAs using performance scores before full training. This preliminary assessment allows the system to identify promising candidates early in the generation process, filtering out poor performers before investing significant computational resources in their full training, thus improving accuracy efficiency.
Data Source
AI summary
Method and system for automatic generation of a target machine learning (ML) pipeline for executing a pre-determined task. The method includes accessing candidate ML pipelines, accessing a plurality of variation operators associated with probability distribution functions, accessing training data and iteratively updating the candidate ML pipelines. An update iteration of a given candidate ML pipelines includes applying, in parallel, one or more of the variation operators to the candidate ML pipeline, training the candidate ML pipeline based on the training data, executing the candidate ML pipeline to determine a performance score, dynamically adjusting the probability distribution functions of the one or more variation operators based on the performance scores and selecting a given variation operator based on the probability distribution functions thereof for a current iteration. The method also includes identifying a given candidate ML pipeline having a highest performance score for executing the pre-determined task.