Decentralized Edge Computing Neural Network Stacked Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional edge computing networks face challenges in decentralized management due to dynamic topologies and heterogeneous device interactions, leading to suboptimal network performance and inefficient data processing.
Innovation Solution
The Platform for Learning Autonomy in Distributed Optimization (PLAyDO) framework employs a stacked architecture of neural networks to enable decentralized edge computing, allowing edge devices to share information and make decisions through multi-hop message propagation, even among devices not directly connected, using message-generating neural networks (MGNNs) and decentralized optimizing neural networks (DONNs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a centralized network structure is used to process all data at a central processor, then data processing can be performed with centralized control, but data processing speed decreases when data concentration occurs and the system lacks scalability
Solution Approach 1:
The patent divides the centralized data processing function into distributed edge computing nodes. Each edge device performs local computation independently, segmenting the processing load across multiple nodes rather than concentrating it at a single central processor. This segmentation enables parallel processing and eliminates the bottleneck of centralized data concentration.
Solution Approach 2:
The patent transitions from a single-point centralized processing architecture to a multi-point distributed processing architecture. By adding the spatial dimension of distribution across multiple edge devices, the system achieves both local processing capability and global coordination, resolving the contradiction between processing speed and centralized capacity.
2Adaptability or versatility
If conventional distributed network optimization methods are used, then decentralized computation can be performed, but complete connection between all devices is required which is not suitable for dynamic edge computing networks
Solution Approach 1:
The patent introduces a mediator mechanism where edge devices share computational results and network state information with their immediate neighbors. This intermediary approach allows indirect information propagation through the network, enabling devices to make optimized decisions without requiring direct connections to all other devices. The mediator mechanism works effectively in dynamic topologies where connections change over time.
3Ease of operation
If decentralized management is implemented in wireless networks with dynamic topologies, then network autonomy and flexibility improve, but coordination among devices with heterogeneous cooperation policies and computation power becomes difficult
Solution Approach 1:
The patent employs parameter changes through iterative optimization. Devices adjust their computation and communication parameters based on feedback from network conditions and neighbor devices. By dynamically changing operational parameters rather than following fixed protocols, devices with heterogeneous capabilities can coordinate effectively. The iterative process allows each device to adapt its behavior to match network demands while maintaining autonomy.
Data Source
AI summary
The present disclosure provides a method and a device for decentralized edge computing. An edge device for implementing decentralized edge computing using a plurality of neural networks having a stacked architecture is provided, including an uncooperative unit configured to generate an initial decision using a neural network of zero-th layer that is a lowest layer of the stacked architecture, an adjacently cooperative unit configured to generate a first transmission messages and a first decision using a pair of neural networks of a first layer, and a network-wide cooperative unit configured to generate a second transmission messages and a second decision using a pair of neural networks of a second layer.