RL for Infrastructure Optimization
Applying RL to auto-scaling decisions in cloud-native AI deployments
Investigating how deep reinforcement learning agents can make proactive auto-scaling decisions for GPU-intensive inference workloads, reducing both cost and latency.
- Hypothesis
- RL agents outperform threshold-based autoscalers for GPU workloads
- Reinforcement Learning · Cloud
- Methodology
- PPO agent trained on 6 months of production traffic data
- Stable Baselines3 · GCP · Prometheus
- Findings
- 40% cost reduction with 15% lower p99 latency
- Deployed in staging environment