An Adaptive Attention-Based GRU Framework with Reinforcement Learning for Cold-Start Prediction and Mitigation in Serverless Computing

Serverless computing, exemplified by AWS Lambda, suffers from cold start latencies that impair performance in dynamic workloads. This study proposes an Adaptive Attention-GRU framework integrated with a Deep Deterministic Policy Gradient reinforcement learning agent to predict and preempt cold starts through dynamic provisioned concurrency adjustments. The Adaptive Attention-GRU leverages hierarchical attention and multi-scale feature encoding to capture short- and long-term patterns in Lambda metrics from AWS CloudWatch, while a dynamic thresholding mechanism triggers pre-warming via Step Functions and PyTorch Lightning with Ray RLlib integration. Evaluated on 90 days of real-world traces across 127 functions, the model achieves 93.63% accuracy, 98.61% ROC-AUC, and 92.46% recall, reducing cold starts by 94.7% with only 23.1% cost increase over baseline scaling. Key innovations include adaptive instantaneous feature weighting, multi-scale temporal modeling, and cost-aware RL autoscaling, offering a validated tool for optimizing latency, cost, and resource efficiency in production serverless environments.