Collaborative safe Bayesian optimization

Important results from the project

The project met its goals by developing Collaborative Safe Bayesian Optimization (CoSBO), extending safe optimization to multi-agent systems for telecommunications. Achievements include the first application of safe Bayesian optimization to mobile network parameter tuning and improved sample efficiency through agent collaboration. CoSBO showed better performance during initial optimization phases while maintaining safety constraints and robustness to weak correlation between collaborators.

Expected long term effects

This project establishes foundations for AI-native network optimization that could influence 6G infrastructure development. Long-term, the collaborative framework may enable better network automation. The research contributes to the broader goal of self-optimizing networks, though practical deployment requires significant additional development. The methodology could extend beyond telecommunications to other safety-critical multi-agent systems.

Approach and implementation

The project started with literature review of safe Bayesian optimization and collaborative methods. SafeOpt-MC was applied to antenna parameter tuning in a network simulator, then extended to create the collaborative algorithm. It was evaluated against baselines with results documented in a final thesis. Knowledge dissemination occurred through presentations at the host company and university. The project executed as planned with weekly supervision and feedback from Ericsson AI experts.