Grouped RL Congestion Control for Unstable Network Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing congestion control algorithms are inflexible and struggle to adapt to changing network conditions, leading to degraded data transmission performance and unstable network performance in new scenarios.
Innovation Solution
A data transmission control method utilizing reinforcement learning (RL) models to analyze historical network performance data, group data transmission objects, and calculate comprehensive network performance data to determine optimized data transmission parameters for each object group, enabling adaptive control based on real-time network conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional congestion control algorithms (CUBIC, BBR) are used, then network transmission quality is improved to some extent, but the system cannot adapt to changing network status and new scenarios, leading to degraded performance
Solution Approach 1:
The patent implements dynamic congestion control by training an RL model offline with extensive network state samples and online interaction data. The model learns to dynamically adjust transmission parameters based on current network conditions, replacing static rule-based algorithms with adaptive intelligence that evolves with changing network status.
Solution Approach 2:
The system changes the control parameters from fixed algorithmic rules to flexible RL model outputs. The trained model selects optimal transmission parameters (bandwidth, RTT, packet intervals) based on learned patterns from historical data, enabling parameter adaptation to various network scenarios rather than relying on predetermined congestion control rules.
2Ease of manufacture
If congestion control algorithms based on expert experience and fixed rules are used, then implementation is straightforward, but the algorithms cannot adapt to new scenarios, leading to unstable network performance
Solution Approach 1:
The system performs preliminary training of the RL model offline using extensive network state samples and online interaction data before deployment. This pre-learning phase captures expert knowledge and network patterns, allowing the system to adapt to new scenarios without requiring complex real-time rule adjustments or manual reconfiguration.
Solution Approach 2:
The patent uses the trained RL model as a copy of expert congestion control knowledge. Instead of implementing complex rule-based algorithms, the system deploys a learned policy that replicates expert decision-making patterns, simplifying implementation while maintaining adaptability to various network conditions through the model's generalization capability.
Data Source
AI summary
A method includes obtaining historical network performance data of each of data transmission objects, calculating, based on the historical network performance data, group network performance data of each of object groups, performing weighted calculation on the group network performance data of the object groups to obtain comprehensive network performance data of the data transmission objects, inputting the comprehensive network performance data and the group network performance data into a trained RL model to obtain a data transmission quality index of each object group, for each object group, determining a data transmission parameter of the object group based on the data transmission quality index and the group network performance data of the object group, and determining, in response to receiving a data transmission request transmitted by a target data transmission object, a target data transmission parameter based on a target object group to which the target data transmission object belongs.