PhD Thesis defense “Model-Based Reinforcement Learning for Dynamic Resource Allocation in Cloud Environments”

Speaker : Thomas Tournaire
Date: 02/05/2022
Time: 10:00 am - 1:00 pm
Location: LINCS Seminars room


The emergence of new technologies (Internet of Things, smart cities, autonomous vehicles, health, industrial automation, …) requires efficient resource allocation to satisfy the demand. These new offers are compatible with new 5G network infrastructure since it can provide low latency and reliability. However, these new needs require high computational power to fulfill the demand, implying more energy consumption in particular in cloud infrastructures and more particularly in data centers. Therefore, it is critical to find new solutions that can satisfy these needs still reducing the power usage of resources in cloud environments. In this thesis we propose and compare new AI solutions (Reinforcement Learning) to orchestrate virtual resources in virtual network environments such that performances are guaranteed and operational costs are minimised. We consider queuing systems as a model for clouds IaaS infrastructures and bring learning methodologies to efficiently allocate the right number of resources for the users. Our objective is to minimise a cost function considering performance costs and operational costs. We go through different types of reinforcement learning algorithms (from model-free to relational model-based) to learn the best policy. Reinforcement learning is concerned with how a software agent ought to take actions in an environment to maximise some cumulative reward. We first develop queuing model of a cloud system with one physical node hosting several virtual resources. On this first part we assume the agent perfectly knows the model (dynamics of the environment and the cost function), giving him the opportunity to perform dynamic programming methods for optimal policy computation. Since the model is known in this part, we also concentrate on the properties of the optimal policies, which are threshold-based and hysteresis-based rules. This allows us to integrate the structural property of the policies into MDP algorithms. After providing a concrete cloud model with exponential arrivals with real intensities and energy data for cloud provider, we compare in this first approach efficiency and time computation of MDP algorithms against heuristics built on top of the queuing Markov Chain stationary distributions. In a second part we consider that the agent does not have access to the model of the environment and concentrate our work with reinforcement learning techniques, especially model based reinforcement learning. We first develop model-based reinforcement learning methods where the agent can re-use its experience replay to update its value function. We also consider MDP online techniques where the autonomous agent approximates environment model to perform dynamic programming. This part is evaluated in a larger network environment with two physical nodes in tandem and we assess convergence time and accuracy of different reinforcement learning methods, mainly model-based techniques versus the state-of-the-art model-free methods (e.g. Q-Learning). The last part focuses on model-based reinforcement learning techniques with relational structure between environment variables. As these tandem networks have structural properties due to their infrastructure shape, we investigate factored and causal approaches built-in reinforcement learning methods to integrate this information. We provide the autonomous agent with a relational knowledge of the environment where it can understand how variables are related to each other. The main goal is to accelerate convergence by: first having a more compact representation with factorisation where we devise a factored MDP online algorithm that we evaluate and compare with model-free and model-based reinforcement learning algorithms; second integrating causal and counterfactual reasoning that can tackle environments with partial observations and unobserved confounders.