Partitioned Training Dataset Update for Wireless Channel Variations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI/ML techniques in wireless communications face challenges in maintaining robustness due to variations in environmental measurements, as existing training datasets are not effectively updated to capture latest channel variations without replacing older measurements or causing dataset growth and contamination.
Innovation Solution
The training dataset is partitioned into multiple dataset groups, each associated with a time stamp and weight, allowing for partial updates by replacing older dataset points with newer ones, updating weights based on correlation, and adding or removing groups, thereby maintaining relevance without fully replacing the dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the training dataset is fully replaced to capture latest channel variations, then the model robustness is improved, but useful older measurements are discarded and dataset growth becomes unbounded
Solution Approach 1:
The training dataset is segmented into multiple dataset groups, where each group contains a subset of training data points. This segmentation allows selective updating of individual groups based on their age and relevance, rather than replacing the entire dataset. Older groups can be retained while newer groups are added or updated, preserving useful historical measurements while incorporating latest channel variations.
Solution Approach 2:
Each dataset group is associated with parameters such as age indicators and weights that track its temporal relevance. By changing these parameters over time and using them to guide update decisions, the system can dynamically determine which groups to update, retain, or remove, thereby maintaining model robustness without unbounded dataset growth.
2Reliability
If the training dataset is fully replaced to capture latest channel variations, then the model robustness is improved, but dataset growth becomes unbounded
Solution Approach 1:
By dividing the dataset into manageable groups, the system can control the size of each group independently. When updating, only necessary groups are modified or added, preventing unbounded growth of the entire dataset while maintaining sufficient diversity for model robustness.
Solution Approach 2:
The system discards outdated or redundant dataset groups while recovering and retaining valuable historical measurements in other groups. This selective discarding and recovery mechanism maintains dataset size within reasonable bounds while preserving essential training information.
3Reliability
If older dataset points are discarded to prevent contamination, then dataset purity is improved, but useful historical measurements are lost
Solution Approach 1:
Segmenting the dataset into age-based groups allows the system to identify and handle potentially contaminated older data separately from newer, cleaner data. Useful historical measurements can be retained in protected groups while obviously contaminated groups are updated or removed, maintaining overall dataset purity without losing valuable information.
Solution Approach 2:
By introducing parameters like age indicators and confidence weights, the system can track the quality and relevance of each dataset group over time. These parameters enable intelligent decisions about which groups to update or remove based on actual contamination levels rather than simply discarding all older data, thus preserving useful historical measurements while maintaining purity.
Data Source
AI summary
Various aspects of the present disclosure relate to training dataset updates. A training dataset is partitioned into multiple dataset groups and each dataset group includes one or more training datapoints. Each dataset group is associated with a first label and a second label. The first label corresponds to a temporal or time-domain related parameter, such as a time stamp or a time duration. The second label is at least one of a weight or a value associated with a characteristic of the dataset. The training dataset is updated based on at least one of the first label or the second label, such as by updating a subset of values of the second label, removing a dataset group, or adding a new dataset group to the training dataset. Updated information corresponding to the updated training dataset can then be sent from one device to another.