ML Pipeline Evolution for Real-Time Data Center Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generating machine learning pipelines and models for large data centers within limited processing resources and time constraints is challenging, as existing methods require extensive processing and memory, which is limiting in real-time operations.
Innovation Solution
A computer-implemented system and method that generates machine learning pipelines and models efficiently by using a multi-objective genetic algorithm to optimize ML pipelines, reducing processing time to about four hours and memory usage by a factor of 2-3, while maintaining similar performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ML pipeline generation methods are used, then model performance can be achieved, but processing time becomes excessively long (30+ hours) and memory usage becomes unmanageable
Solution Approach 1:
The patent segments the ML pipeline generation process into multiple generations, where each generation processes a subset of pipelines with increasing data volume. This progressive segmentation allows the system to evaluate pipelines incrementally, reducing the time penalty associated with evaluating all pipelines on the full dataset while maintaining performance through iterative refinement.
Solution Approach 2:
The patent performs preliminary filtering in early generations by evaluating pipelines on smaller data subsets before progressing to larger datasets. This preliminary action eliminates poorly performing pipelines early in the process, preventing wasted computation on ineffective models and significantly reducing total processing time while preserving the potential for high-performance solutions.
2Measurement precision
If conventional ML pipeline generation methods are used, then model performance can be achieved, but memory usage increases by a factor of 2-3 times
Solution Approach 1:
The patent divides the dataset into multiple subsets and processes them across different generations. Each generation loads only the necessary data subset into memory, avoiding the need to load the entire dataset at once. This segmentation reduces peak memory usage by a factor of 2-3 while still enabling comprehensive pipeline evaluation through iterative processing.
Solution Approach 2:
The patent evaluates a limited subset of pipelines in each generation rather than all pipelines simultaneously. This partial action approach reduces memory requirements by processing pipelines incrementally, while the evolutionary framework ensures that sufficient pipelines are evaluated to identify high-performance solutions without requiring excessive memory resources.
3Measurement precision
If the entire dataset is used for pipeline evaluation, then model performance can be optimized, but processing time increases significantly
Solution Approach 1:
The patent performs preliminary evaluations on smaller data subsets in early generations to quickly identify and eliminate poor-performing pipelines. This preliminary action provides a fast initial filtering mechanism that reduces the number of pipelines requiring full-dataset evaluation, thereby improving overall productivity while maintaining the ability to discover high-performance models.
Solution Approach 2:
The patent dynamically adjusts the data volume used for evaluation across generations, starting with smaller subsets and progressively increasing to larger subsets. This dynamic approach allows the system to balance performance optimization and processing speed adaptively, spending less time on early filtering and more time on refined optimization of promising pipelines.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and a method for operating a data center. The operating comprising executing predictive maintenance of the data center or network monitoring of the data center. The operating being based on a generated machine learning (ML) pipeline, the method comprising accessing data relating to operations of the data center, the data being suitable for evaluating respective performances of the plurality of ML pipelines. The method comprises generating the plurality of ML pipelines, selecting a sub-set of ML pipelines from the plurality of ML pipelines, evolving the sub-set of ML pipelines to generate evolved ML pipelines, selecting a sub-set of evolved ML pipelines from the evolved ML pipelines and iterating until determination is made that iterating is to be stopped. The method also involves operating, by an operation monitoring system of the data center, at least one of the ML pipelines from the sub-set of evolved ML pipelines.