Federated Growing Neural Gas Model for Radio Access Network Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies face challenges in optimizing the usage of network resources, particularly in wireless communication networks, where local growing neural gas models represent local state models of radio access network nodes but lack a method to merge these models into a global state model efficiently.
Innovation Solution
The proposed solution involves an apparatus comprising at least one processor and memory that receives local growing neural gas models from distributed trainers, trains a global growing neural gas model based on these local models, and represents a global state model of multiple radio access network nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local growing neural gas models are trained independently at distributed trainers, then local state model accuracy is improved, but the ability to represent global network state deteriorates
Solution Approach 1:
The patent merges multiple local growing neural gas models from distributed trainers into a single global model. The central trainer receives local models, extracts unit vectors and weights, and combines them to form a global state model that represents the entire network while preserving local accuracy characteristics.
Solution Approach 2:
The central trainer acts as an intermediary between distributed local trainers and the global model. It receives local models, processes them by extracting and combining unit vectors, and generates the global state model, enabling coordination between local precision and global representation.
2Adaptability or versatility
If a global state model is created by collecting all local training data centrally, then global network state representation is improved, but network resource usage efficiency deteriorates
Solution Approach 1:
Instead of transferring all local training data to the central trainer, the patent extracts only the essential model parameters (unit vectors and weights) from each local growing neural gas model. This extraction approach creates the global model while minimizing data transmission and preserving energy efficiency.
Solution Approach 2:
The patent uses copies of local model parameters (unit vectors and weights) to construct the global model. Rather than moving original data, the central trainer combines parameter copies from distributed models, reducing network resource consumption while achieving global representation.
3Loss of information
If local models are merged into a global model by collecting all parameters, then global state model completeness is improved, but model complexity and processing requirements deteriorate
Solution Approach 1:
The patent segments the global model construction process into distinct steps: receiving local models, extracting unit vectors, identifying winning units, and combining weights. This segmentation manages complexity by breaking down the merging process into manageable operations rather than handling all parameters simultaneously.
Solution Approach 2:
The patent applies local quality by treating different components of the model merging process differently. Unit vectors are extracted and normalized, winning units are identified through comparison, and weights are combined based on their significance. Each component receives appropriate processing to maintain completeness while managing overall complexity.
Data Source
AI summary
Disclosed is a method comprising receiving a plurality of local growing neural gas models from a plurality of distributed trainers, wherein a local growing neural gas model of the plurality of local growing neural gas models represents a local state model of at least one radio access network node; and training a global growing neural gas model based on the plurality of local growing neural gas models, wherein the global growing neural gas model represents a global state model of a plurality of radio access network nodes.