Reinforcement Learning-Based Adaptive Utilization of Resources in Cloud Networks

The necessity of dynamic resource allocation in cloud systems has opened door for reinforcement learning based approaches in the domain. Cloud computing services provide users with on-demand resources for various workloads with distinct service performance requirements. However, dynamic workloads, varying resource requirements, and the tension between efficient performance and cost efficiency can place significant strain on the resource management within these types of platforms. Because cloud systems are inherently dynamic processes, traditional resource allocation approaches, like manual provisioning or using predefined optimization rules, are not likely to solve the problem satisfactorily which makes adaptive approaches (like those driven with RL) as appealing alternatives. A RL-based resource allocation framework includes an agent that interact with the cloud environment by trying different actions and learns the optimal policies as a result. This usually involves specifying a reward function, where the agent receives a signal indicating how well the system is performing, along with a number of criteria like minimizing latency, maximizing throughput or cutting operational costs. The RL agent works in a dynamical manner, meaning it adapts its resource allocation decision rules optimally based on previously made actions so as to improve efficiency in future decisions. This is fairly effective in managing cloud(such) environments which are extremely dynamic, i.e. resource requirement of these systems alters with time due varying user demands, system conditions and external dependencies. We propose the use of RL to provide dynamic resource allocation in cloud systems which brings us to discuss certain critical problems in this paper. The first part introduces the main elements of an RL framework for resource management agent: the environment, the state space, the action space, and the reward function. Cloud system state space may refer to cloud system present configuration abundance of resources, workload requirements, and performance metrics. The possible action space captures resource allocation decisions, such as adding or removing virtual machines, redistributing storage or modifying the network capacity.