Pseudo Gradient Teaching Learning Based Optimization for Proactive SLA Prediction in 6G Networks

Proactive prediction of Service Level Agreement (SLA) satisfaction lets 6G radio access networks act before a link misses its throughput and latency targets. This paper proposes Pseudo Gradient Teaching Learning Based Optimization (PGTLBO), which adds to the teacher step the direction from the worse half to the better half of the ranked population, and uses it to train a multilayer perceptron. All trainers receive the same number of function evaluations, and a random direction control isolates the effect of the direction. On 6G SLA prediction for unseen clients, PGTLBO reaches 0.846 accuracy, tied with standard TLBO (0.844), and outperforms the genetic algorithm, particle swarm optimization, and Harris hawks optimization on all 20 splits. PGTLBO reaches a lower training loss than TLBO in 57 of 80 paired runs and its direction beats a random one, but test accuracy is limited by the data.