Regional Agent Network Optimization via Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wireless network self-optimization algorithms, both conventional and artificial intelligence-based, face challenges in adapting to the unique conditions of individual base stations due to reliance on fixed rules and single-agent models, leading to ineffective self-optimization and convergence issues in multi-agent environments.
Innovation Solution
A network optimization method that models problems in specific regions to create multiple agents using reinforcement learning, allowing for adaptive training and optimization based on geographic and performance data, enabling each agent to make tailored adjustments within its environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional network self-optimization algorithms use manually developed rule tables, then the self-optimization process is simple to implement, but the optimization effectiveness is poor because the same rules cannot adapt to different base station conditions
Solution Approach 1:
The patent divides the network into multiple regions and creates separate agents for each region. Each agent is trained independently using reinforcement learning to handle local optimization tasks, allowing different base stations to have customized optimization strategies tailored to their specific conditions rather than applying uniform rules across the entire network.
Solution Approach 2:
The patent uses reinforcement learning to dynamically adjust optimization parameters based on real-time network conditions and base station characteristics. This allows the system to adapt parameters such as handover thresholds, resource allocation, and power settings according to local conditions, transforming static rule-based parameters into dynamic, context-aware values.
2Device complexity
If artificial intelligence-based algorithms use single-agent models, then the training process is manageable, but the convergence problem occurs when deploying to large-scale multi-agent networks
Solution Approach 1:
The patent segments the large-scale network into multiple smaller regional regions, each managed by its own agent. This segmentation reduces the complexity of training for each individual agent while ensuring that when deployed across the entire network, each agent can converge reliably within its local context without the instability that arises in monolithic multi-agent systems.
Solution Approach 2:
Each agent is trained to handle the specific characteristics and conditions of its local region, making the optimization strategy locally optimized rather than globally uniform. This local quality approach ensures that each agent develops expertise for its specific environment, improving convergence reliability while maintaining manageable training complexity.
3Reliability
If network optimization is performed network-wide, then comprehensive coverage is achieved, but the computational overhead is excessive
Solution Approach 1:
The patent divides the network into multiple independent regions with separate agents for optimization. This segmentation allows computational resources to be distributed and focused on local regions rather than processing the entire network simultaneously, significantly reducing the computational overhead while maintaining comprehensive coverage through the collective action of multiple regional agents.
Solution Approach 2:
Instead of optimizing the entire network at once, the patent implements partial optimization by focusing each agent on its specific regional subset. This partial action approach reduces computational overhead by limiting the scope of each optimization process to manageable regional boundaries while still achieving network-wide optimization effects through the aggregation of regional improvements.
Data Source
AI summary
A network optimization method, a device, and a non-transitory computer-readable storage medium are disclosed. The method may include: modeling problems existing in cells in a first region to obtain N agents, a modeling method and a training method, where a proportion of cells in which existing problems belong to a same problem category among the cells contained in the first region is greater than or equal to a preset threshold, geographic locations of the cells contained are consecutive, and an outline of the first region is an outwardly convex figure, and N is an integer greater than or equal to 1; and for each of the agents, determining an initial model of the agent according to the modeling method and the training method, and training the initial model of the agent using a reinforcement learning method according to the modeling method and the training method.