Offline Meta-Reinforcement Learning for Cold-Start Adaptation in Multi-Cloud Auto-Scaling, Addressing Data Scarcity and Action-Space Explosion in CARL Frameworks

• Billy Elly

  Lautech

  Author

Multi-cloud auto-scaling faces persistent challenges in cold-start scenarios where historical data is scarce and the joint action space across providers grows exponentially. The Cost-Aware Reinforcement Learning (CARL) framework addresses SLA-aware multi-cloud scaling but assumes sufficient offline data and tractable action spaces. This study develops an Offline Meta-Reinforcement Learning (OMRL) approach to cold-start adaptation in CARL-based multi-cloud auto-scaling. Using a design-based research methodology with retrospective analysis of multi-cloud workload traces and prospective simulation, we train a context-based OMRL agent with task inference mechanisms to generalize across heterogeneous cloud environments. Results demonstrate that the proposed framework achieves 89.4% accuracy in predicting optimal scaling actions during cold-start periods, representing a 17.2 percentage-point improvement over standard CARL baselines. The method reduces SLA violations by 23.6% while maintaining cost efficiency within 2.1% of oracle performance. Key contributions include a task-representation learning mechanism for multi-cloud heterogeneity and an action-space factorization strategy that reduces effective decision dimensionality by 68%. The framework provides cloud administrators with a practical tool for rapid adaptation in new deployment contexts without extensive historical data collection.

• pdf

Copyright (c) 2026 Billy Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.