3GPP TR 37.817 explores the use of AI/ML techniques to optimize mobility management—particularly handover decisions—to mitigate issues like ping-pong handovers and cell overload by leveraging UE/network measurements and historical data.
I'm interested in simulating this use case in NetSim, specifically:
- How can AI/ML-based handover parameter tuning be modeled within the NetSim framework?
- What’s the best way to collect and structure training data from simulations (e.g., UE signal strength, handover success/failure, mobility patterns)?
- How can one integrate Python-based ML models into the simulation loop for real-time decision-making or periodic policy updates?
Has anyone attempted this or come across a relevant example? I’d appreciate any insights, shared experiences, or even experimental setups to get started.
These two use cases may help:
1. https://tetcos.com/5G-Load-Balancing.html - Use of Cell specific offset for load balancing
2. https://tetcos.com/network-slicing.html - Use of network slicing to efficiently allocate RAN resources
The current model shows how to adjust CIO to achieve load balancing by monitoring PRB utilization. A positive CIO makes an underloaded cell more attractive, while a negative CIO makes an overloaded cell less attractive.
This adjustment needs to be run in a loop: modify CIO, observe PRB utilization, and repeat. Over multiple iterations, this provides a mechanism for convergence toward balanced load across cells.
To automate this process, RL can be applied. However, the MDP must be defined carefully to reflect the network dynamics and objectives.
To implement AI/ML-assisted load balancing from 3GPP TR 37.817 in NetSim, you could perhaps model the problem as an RL environment:
States: (For example) PRB utilization, SINR 5%, SINR mean, average throughput, UE count (continuous)
Actions: Adjust CIO of each cell by +1 dB or −1 dB
Reward: Neagtive of (Variance of PRB utilization across cells plus penalty for excessive handovers)
Algorithm: Use PPO for continuous state input and discrete actions
NetSim simulates scenarios for each state, action pair and feedbacks the reward feedback. Transitions can be based on random mobility.
Thank you all,
Should the goal be to equalize the load or should the objective be overall performance optimization? Does it not make sense to distribute load across the gNBs so as to maximize performance?
Load balancing fairness should be tunable. The objective for any operator is to maximize revenue.
Yes, the overall goal of load balancing is to optimize performance. Since maximizing raw throughput would bias toward near-cell users, the commonly used metric is sum-log throughput. If Ri denotes the throughput of UE i, the objective is to maximize the sum of log(Ri) over all UEs.
This formulation balances throughput with fairness across near-cell, mid-cell, and cell-edge users.
Hence, the load balancing problem can be to maximize the sum-log throughput, subject to resource and association constraints.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Energy