A Call Admission Control Scheme Using NeuroEvolution Algorithm in Cellular Networks

Xu Yang, John BighamIJCAI 2007

Cellular networks operate under strict bandwidth constraints. As mobile traffic grows more complex with multimedia services, managing the trade-off between admitting new users and maintaining existing connections becomes a critical challenge. This paper, originally presented at IJCAI 2007, introduces a novel Call Admission Control (CAC) scheme that leverages the NeuroEvolution of Augmenting Topologies (NEAT) algorithm to optimize network performance in dynamic environments.

The Challenge of Call Admission Control

In a cellular network, two primary metrics define the quality of service:

  1. Call Blocking Rate (CBR): The frequency at which new call requests are rejected due to lack of capacity.
  2. Call Dropping Rate (CDR): The frequency at which active calls are terminated when a user moves into a full cell (handoff).

Dropping an active call is significantly more disruptive to the user experience than blocking a new one. Consequently, networks often reserve "guard channels" for handoffs. However, determining the optimal reservation policy is difficult in modern networks with multi-class traffic (voice, video, data) and fluctuating demand. Traditional Reinforcement Learning (RL) approaches often struggle with the "state space explosion"—the exponential increase in possible call combinations—making them difficult to scale for real-world deployment.

The CN Architecture: NeuroEvolution for CAC

The authors propose the CN (Call Admission Control using NeuroEvolution) scheme. Unlike standard neural networks that use fixed architectures and backpropagation, CN utilizes the NEAT algorithm to evolve both the weights and the topology of the neural network.

The CN architecture utilizing the NEAT algorithm with internal and external supervisors.

As shown in the architecture diagram, the system processes several real-time inputs, including current CBR/CDR levels, the number of ongoing calls per class, and the nature of the new user request. To ensure operational safety during the evolutionary process, two supervisors are employed:

  • Internal Supervisor: Prunes policies that exhibit irrational behavior, such as rejecting all incoming traffic.
  • External Supervisor: Acts as a hard constraint, preventing the system from attempting physically impossible actions, such as accepting a call when the cell is at maximum capacity.

Experimental Methodology

The scheme was evaluated using a custom discrete-event network traffic simulator. The environment modeled a single cell with 20 bandwidth units and two traffic classes (C1 and C2) requiring 1 and 2 units, respectively. The researchers tested the system under two primary conditions:

  • Constant Traffic Loads: To establish a baseline against traditional RL-CAC methods.
  • Time-Varying Loads: Using a sine function to simulate the realistic "ebb and flow" of daily urban traffic patterns.

Performance in Dynamic Environments

A key finding of the study is the system's ability to "sense" traffic trends. By feeding real-time CBR and CDR back into the network, the evolved controllers could adjust their admission policies on the fly without requiring a complex mathematical model of the traffic arrival process.

CBR and CDR performance across evaluation intervals under time-varying traffic loads.

In time-varying scenarios, the CN scheme demonstrated superior adaptability. While traditional RL schemes often lag when traffic patterns shift, the CN scheme successfully maintained the handoff dropping rate (CDR) below target thresholds (e.g., 1% or 2%) while maximizing service for new calls. The evolutionary approach allowed the system to discover efficient, non-standard network topologies that would be nearly impossible to design manually.

Conclusion

The CN scheme proves that NeuroEvolution is a viable and robust alternative to traditional reinforcement learning for complex network management tasks. By evolving both structure and weights, the system achieves near-optimal performance in stable conditions and superior flexibility in non-stationary environments. Future work may expand this model to multi-cell environments and adaptive services that utilize variable bitrates.

Summary updated

BibTeX
@inproceedings{yang2007call,
  author = {Xu Yang and John Bigham},
  title = {{A Call Admission Control Scheme Using NeuroEvolution Algorithm in Cellular Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure