Deep Reinforcement Learning for Wireless Network Cell Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional wireless network optimization methods face challenges in achieving lower optimization costs, shorter optimization times, and higher accuracy as networks become more complex, requiring innovative approaches to adjust cell parameters efficiently.
Innovation Solution
The method employs deep reinforcement learning (DRL) to train a neural network that adjusts cell parameters in wireless networks by generating experience tuples based on state transitions and reward values, using both DRL-generated and expert-generated experiences to optimize cell settings and update weight values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional wireless network optimization methods are used, then network performance can be improved, but optimization time and cost increase significantly as networks become more complex
Solution Approach 1:
The system implements self-service through autonomous DRL agents that automatically optimize cell parameters without human intervention. The neural network learns optimal configurations through reinforcement learning from environmental feedback, enabling the network to self-optimize in real-time, thereby reducing optimization time while maintaining performance improvements
Solution Approach 2:
The patent replaces traditional mechanical/optical measurement systems with electronic signal processing and machine learning algorithms. Instead of physical drive tests and manual measurements, the system uses electronic data collection from UEs and processing through neural networks to determine optimal parameters, significantly accelerating the optimization process
2Reliability
If conventional wireless network optimization methods are used, then network performance can be improved, but optimization cost increases significantly as networks become more complex
Solution Approach 1:
The autonomous DRL-based system eliminates the need for expensive human expert intervention by implementing self-service optimization. The neural network agents automatically learn and apply optimization strategies, reducing labor costs and making the optimization process more economically viable for complex networks
Solution Approach 2:
The DRL framework provides a universal optimization approach that can handle multiple network parameters and scenarios through a single learned model. The neural network learns generalizable patterns that apply across different network conditions, reducing the need for multiple specialized optimization tools and expertise
3Loss of time
If deep reinforcement learning is used for wireless network optimization, then optimization time and cost are reduced, but the system requires generating and processing large amounts of experience tuples
Solution Approach 1:
The system performs preliminary action by pre-collecting and storing experience tuples in a replay buffer during network operation. This allows the DRL agent to learn from historical data without requiring real-time data generation during optimization, reducing the immediate computational burden while maintaining learning effectiveness
Solution Approach 2:
The patent applies partial action by selectively sampling experience tuples from the replay buffer rather than processing all available data. The DRL agent uses experience replay with batch sampling to learn from representative subsets of data, reducing processing requirements while maintaining learning quality
Data Source
AI summary
A neural network is trained using deep reinforcement learning (DRL) techniques for adjusting cell parameters of a wireless network by generating a plurality of experience tuples, and updating the neural network based on the generated experience tuples. The trained neural network may be used to select actions to adjust the cell parameters. Each experience tuple includes a cell identifier, a first state, a second state, an action applied to the cell that moves the cell from the first state to the second state, a local reward, and a global reward. The neural network is updated based on whether or not each action is acceptable, which is determined based on the global reward and the local reward associated with each action.