Reinforcement Learning Based Adaptive Task Scheduling in Cloud Environments

Abstract

Efficient task scheduling is critical for optimizing resource utilization and minimizing response times in cloud computing environments. Traditional task scheduling algorithms rely on static heuristics that limit their adaptability to dynamic workloads and varying resource demands. This paper presents a novel reinforcement learning (RL) based approach that employs a Double Deep Q-Network (DDQN) for intelligent task allocation in cloud infrastructure. The RL agent operates within a set state space, capturing real-time information about CPU availability, current task loads, and task characteristics across a cloud environment. The model was trained within CloudSim and validation performed at set intervals. Experimental results demonstrate that the model outperformed traditional algorithms across varying workload conditions. To evaluate transferability and robustness, we conducted additional experiments by deploying the trained agent in a resource-constrained environment. With a 20 % reduction in available resources and minimal finetuning of the pre-trained model, we observed similar performance demonstrating strong adaptability in fault-prone or resourcelimited scenarios. These findings establish deep reinforcement learning as a promising, scalable alternative to conventional task scheduling strategies for modern cloud computing systems.

Department(s)

Computer Science

Document Type

Conference Proceeding

DOI

10.1109/ICCT67417.2025.11374147

Keywords

Cloud Computing, Double Deep Q-Network, Dynamic Load Balancing, Reinforcement Learning, Task Scheduling

Publication Date

1-1-2025

Journal Title

International Conference on Communication Technology Proceedings ICCT

Link to Full Text

Share

COinS