Multilevel Neural Network Equalizer Activation Function Configuration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional neural networks with medium size fail to meet the needs of various high-speed access networks, particularly in flexible PON systems, and have inadequate Bit Error Rate (BER) performance, necessitating a solution to adapt neural network equalizers while controlling their scale within a reasonable range.
Innovation Solution
A method where a second communication device determines and sends configuration parameters to a first communication device to configure a multilevel neural network equalizer, selecting an appropriate activation function class based on channel conditions, allowing the equalizer to better adapt to high-speed access networks and improve system performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional neural networks with medium size are used for equalization, then the equalizer can provide basic compensation for transmission impairments, but the Bit Error Rate (BER) performance is inadequate and the network scale cannot adapt to various high-speed access networks
Solution Approach 1:
The patent implements dynamic adaptation of neural network equalizers by allowing configuration of different activation function classes (e.g., ReLU, tanh, sigmoid, custom activation functions) based on channel conditions and network requirements. This dynamic configuration enables the equalizer to adapt to various high-speed access networks while maintaining adequate BER performance.
Solution Approach 2:
The patent changes key parameters of the neural network equalizer, specifically the activation function class, to optimize performance. By allowing selection from multiple activation function classes and adjusting network scale parameters, the system achieves both adequate BER performance and adaptability to different high-speed access network scenarios.
2Device complexity
If the neural network scale is reduced to control within a reasonable range, then the system becomes more manageable and efficient, but the equalization performance may be compromised
Solution Approach 1:
The patent optimizes the neural network scale by changing parameters such as the number of neurons, layers, and activation function types. This allows the network to maintain an appropriate scale for efficient management while achieving adequate equalization performance through intelligent parameter selection rather than simply increasing network size.
Solution Approach 2:
The patent employs pre-trained neural network models that can be copied and deployed with different configuration parameters. This allows the system to use a standardized base model with varying activation function classes and scale parameters, maintaining manageability while ensuring sufficient equalization performance through parameter adaptation rather than training larger networks from scratch.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to device, method, apparatus and computer-readable storage medium for communications. Multilevel neural network based signal equalization method and apparatus are provided in the communication system. In the method, the second communication device determines and sends to the first communication device the configuration parameters to be used by a first communication device for configuring an equalizer at the first communication device, wherein the configuration parameters indicate one activation function class of a plurality of activation function classes for configuring the equalizer. The method comprises configuring, by the first communication device, the equalizer at the first communication device with the configuration parameters. In this way, the neural network based equalizer can better adapt to the high-speed access networks. Moreover, the first communication can implement the multilevel neural network equalizer based on the configuration parameters indicating the activation function class. Thus, the scale of the neural network may be controlled within a reasonable range and the system performance may be further boosted.