DRL Agent for Self-Organizing Network Cell Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The optimization of cellular mobile communications networks, particularly in 5G networks, faces a complex NP-hard combinatorial optimization problem due to the large number of possible configurations for tunable parameters like antenna tilts, transmission power, and beamforming techniques, making it difficult to systematically explore all solutions.
Innovation Solution
A data-driven Deep Reinforcement Learning (DRL) method and system are employed for Coverage and Capacity Optimization (CCO) in self-organizing networks. This involves creating a simulated network environment using MDT data, network performance indicators, and electromagnetic simulations to train a DQN agent. The agent interacts with the environment to adjust modifiable parameters of network cells, receiving rewards based on network performance metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If all possible configurations of network cell parameters are systematically explored to find optimal Coverage and Capacity, then the completeness of solution search is improved, but the computational complexity and time required become prohibitively large
Solution Approach 1:
The patent segments the vast solution space by dividing it into promising regions and less promising regions based on initial evaluation. The optimization algorithm focuses computational resources on exploring the promising regions in detail, while using coarser exploration for less promising regions. This segmentation allows the system to achieve high optimization precision without requiring exhaustive search of all possible configurations.
Solution Approach 2:
The patent applies local quality by using different exploration strategies for different regions of the solution space. Promising regions receive intensive local search with fine-grained parameter adjustments, while other regions receive broader exploration. This localized approach to optimization quality enables the system to achieve high precision where it matters most without the computational cost of uniform high-precision search everywhere.
2Adaptability or versatility
If the number of tunable parameters per network cell is increased to improve Coverage and Capacity optimization capability, then the optimization potential is improved, but the size of the solution space grows exponentially
Solution Approach 1:
The patent performs preliminary action by first evaluating all possible parameter configurations to identify and mark the promising regions before the main optimization process begins. This preliminary evaluation creates a guide map that directs the subsequent optimization algorithm to focus on regions with high potential, avoiding exhaustive search of the entire exponentially growing solution space. The preliminary action transforms the unmanageable full search problem into a directed search problem.
Solution Approach 2:
The patent applies partial action by deliberately limiting the search to only the most promising regions of the solution space rather than exploring all possible configurations. The system accepts that not all regions will be exhaustively searched, but this partial exploration is sufficient to find optimal or near-optimal solutions. This approach makes the optimization feasible despite the exponential growth of the solution space with increased parameters.
Data Source
AI summary
A method for adjusting network cell parameters in a self-organizing cellular mobile network using a data processing system. The method involves creating an Environment that simulates the network based on radio measurement data, network performance data, and electromagnetic field simulation. A Deep Reinforcement Learning (DRL) Agent interacts with this Environment to simulate the effects of parameter changes on network performance. The Environment calculates and returns a Reward to the DRL Agent, which is used to train the Agent and estimate Q-values. The DRL Agent selects actions based on a policy that balances greedy actions, random actions, and constrained random actions. If the DRL Agent violates a predetermined constraint, the Environment returns a penalizing Reward.