DRL algorithms struggle to outperform calibrated baselines in resource control benchmarks

A new benchmark study, RLScale-Bench, has been developed to evaluate deep reinforcement learning (DRL) algorithms for adaptive resource control. The research found that a properly calibrated rule-based autoscaler often outperforms mainstream DRL algorithms in terms of cost across various workloads. However, DRL algorithms showed promise in handling bursty and flash traffic. The study also highlighted that discrete-action algorithms were significantly better than continuous-action ones in avoiding constraint violations, and no single DRL algorithm proved dominant across all tested workloads. AI

IMPACT Highlights the need for better calibration and reward engineering in DRL for resource control, suggesting current methods may not surpass simpler baselines.

RANK_REASON The cluster contains an academic paper detailing a new benchmark and evaluation protocol for DRL algorithms. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Deep Reinforcement Learning
• Kubernetes Horizontal Pod Autoscaling
• RLScale-Bench
• TD3

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →