Overview

Mobility load balancing is a 3GPP Release 17 AI/ML for NG RAN Use Case. It involves transferring load from overloaded cells to under-loaded neighboring cells, for optimizing network performance and user experience.


Concept

  • Default Association: In NetSim, the default user equipment (UE) association is based on Maximum signal strength (SS-RSRP)
  • Load Balancing Goal: Modify the association/Handover criteria to distribute network load efficiently across available cells.


Load Balancing Algorithm

Inputs

  • Number of RRC connected UEs at each gNB
  • DL and UL CQIs of each UE
  • Time averaged PRB utilization (DL and UL) at each gNB

Possible Outputs

  • UE to gNB associations
  • Handover offsets
  • Cell Individual Offsets

NetSim provides a flexible framework for users to develop and test load balancing algorithms:

  • NetSim passes 'measurements' to the user algorithm
  • The algorithm processes data and returns 'controls or actions'
  • NetSim adjusts the simulation based on algorithm output
  • NetSim then provides performance metrics (KPIs) back to the algorithm
  • These steps occur in a continuous loop, allowing for run-time adjustments.

User Algorithm Development

  • Algorithms can be developed in high-level languages like Python
  • No need for deep knowledge of NetSim internals


Example Scenario

  • 7-cell hexagonal layout
  • 3 sectors per cell, 2 carriers per sector (total 7*3*2 = 42 gNBs in NetSim)
  • 50 active UEs per sector

Additional Considerations

To create more sophisticated load balancing solutions, users could consider the following:

  • PRB utilization between GBR (Guaranteed Bit Rate) and Non-GBR users
  • Time-varying network traffic patterns
  • UE mobility 
  • Minimum throughput per user 

Advanced: An outline for applying Reinforcement Learning (RL) for load balancing

  • Now, let us denote c_ij as the instantaneous rate of a UE and is theoretically a log function of SINR 

  • And let R_ij be the long-term rate and y_ij is the fraction of resources allocated, to UE_i by BS_j 


  • Note that the max RSS association does not balance the load between BSs. The load balancing problem can be solved from optimization theory for a fixed topology
  • Now let's say the SINR changes with time due to user mobility
  • Then RL can be used to decide a "load-aware" UE-BS association i.e., the association is not based on max RSS
  • We explore, Markov decision process/Q-learning based (model-free) RL
    • At state s_t  RL agent selects action a_t by following policy π and receives reward r(s_t, a_t). 
    • The MDP has value function V^π (s), and action value function Q^π (s, a) where α (0≤α≤1) is the discount factor 
    • We assume that the update interval (epoch) ≫ LTE/5G frame length of 10ms
  • State: UE SINRs (γ_1,…, γ_N  ), based on the current association at time t 
  • Action: 
    • Association x_ij (indicator variable showing association of UEi_ to BS_j)
    • Resource allocation y_ij (equals 1/sum_j x_ij, i.e., the reciprocal of the number of UEs associated with a BS. This is exact for Round robin, and on average for PFS)



Power control example involving Reinforcement Learning