IoT Task Allocation Using Performance Prediction for Edge Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing edge deep learning systems lack effective strategies for task scheduling and resource optimization, particularly in scenarios where cloud computing is inadequate due to sensitivity to delay, reliability, privacy, and limited resources on single IoT devices.
Innovation Solution
A method and apparatus for task allocation and network training on IoT devices, involving the construction of computation and resource graphs to generate task allocation strategies, predict performance using a trained network, and optimize resource utilization through intelligent task scheduling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If cloud computing is used to process computation-intensive deep learning tasks, then computing power and storage resources are improved, but delay sensitivity, reliability, and privacy protection deteriorate
Solution Approach 1:
The patent segments the deep learning model into multiple layers and distributes them across different edge devices in the network. Instead of centralized cloud processing, the computation is divided into smaller units that can be executed locally on edge devices, reducing transmission delay and improving reliability for time-sensitive applications.
Solution Approach 2:
The patent transitions from a single-dimensional cloud computing architecture to a multi-dimensional distributed edge computing architecture. By adding the spatial dimension of multiple edge devices and introducing temporal dimensions through iterative optimization, the system achieves both high computing power and low latency simultaneously.
2Productivity
If distributed edge computing is implemented across multiple heterogeneous IoT devices, then resource utilization and reliability are improved, but task allocation complexity increases
Solution Approach 1:
The patent implements a feedback mechanism where the system predicts the performance of different task allocation strategies using a trained neural network, then uses this prediction to guide the actual task allocation. The system continuously learns from the difference between predicted and actual performance, optimizing the allocation strategy iteratively and reducing complexity through experience.
Solution Approach 2:
The patent performs preliminary training of the performance prediction network using historical task allocation data and performance metrics. This pre-trained model can quickly predict the outcome of different allocation strategies without requiring complex real-time calculations, simplifying the task allocation process while maintaining high resource utilization.
3Ease of operation
If hierarchical scheduling algorithm based on model segmentation is used, then task allocation is simplified, but optimization effectiveness deteriorates due to lack of underlying algorithm allocation
Solution Approach 1:
The patent introduces a performance prediction network as an intermediary between the task allocation decision and the actual execution. This neural network mediator evaluates multiple allocation strategies and predicts their performance, enabling the system to make optimized decisions without requiring complex manual configuration or rule-based algorithms.
Solution Approach 2:
The system performs self-optimization by automatically learning the relationship between task allocation strategies and performance metrics through the trained prediction network. Instead of relying on pre-configured hierarchical rules, the system adapts to the specific characteristics of the network and tasks, achieving superior optimization effectiveness while maintaining operational simplicity.
Data Source
AI summary
Disclosed are a task allocation method and apparatus based on an Internet-of-Things device, and a network training method and apparatus. The network training method comprises: determining a training data set; and training a first network on the basis of the training data set, wherein the training data set comprises at least one task allocation strategy and a corresponding actual performance, an actual performance is obtained on the basis of actual execution of a corresponding task allocation strategy, and the first network is used for predicting the performance of a task allocation strategy.