Optimization approaches in distributed systems

The necessity of performing multiple complex tasks in short time in a typical Industry 4.0 environment has enlightened the need to improve computational and storage infrastructures and ensure efficient communication between machines in modern manufacturing plants. Most of these processes are pure digital and can be delivered directly in the Cloud, thus ensuring more resilience and performance. When Cloud resources are delivered in distributed fashion, the global computational effort is spread across all the network, whereas when the process is also decentralized there’s not a single point of decision in system behaviour. As a result, the innate features of distributed and decentralized systems contribute to decreasing the execution time of the tasks and preventing single points of failure in the infrastructure. Within the scope of these systems, a great variety of optimization problems can be implemented. Some of these problems pertain to classical linear or nonlinear approaches, whereas others to Machine Learning techniques, such as Artificial Neural Networks and Deep Reinforcement Learning (DRL), or Swarm Intelligence algorithms. The main research direction of this thesis is the implementation of some common optimization problems in distributed systems. In detail, in part I, the application of the task assignment problem in Blockchain-based environments is proposed in a typical manufacturing environment. Two different original architectures, the first in Ethereum and the second in Hyperledger Fabric, are described with the scope of consuming multiple complex digital processes, such as data mining problems, heavy file format conversion, and 3D rendering, in a parallel fashion in multi agent environments. In both proposed architectures, task delivery is orchestrated by a Smart Contract. The main goal is to find the agent that executes the required task in least time. To this end, a task runtime prediction algorithm is designed by the means of an Artificial Neural Network in the first case and a Deep Reinforcement Learning algorithm in the second case. While in the first case, the main contribution is the combined use of Blockchain with some popular cloud technologies, such as Docker containers and Cloud Storage, the main advantages of the second implementation are that the Smart Contract implements an auction and bidding scheme to deliver tasks, and that the agents learn how to make predictions as they collect new experiences. In addition, a third Blockchain-based architecture is proposed in the context of energy management in which a penalty-reward optimization scheme for the users in a district is implemented based on their consumption. Still in the context of manufacturing, a general model for mass production systems based on Timed Coloured Petri Net and Particle Swarm Intelligence is proposed in part II. In the model, the Flexible Job Shop Sequence Problem is approached by the means of a Particle Swarm Optimization algorithm that can be efficiently implemented in a distributed environment. Finally, in part III, a cooperative and distributed-oriented multi-agent DRL scheme is proposed in the domain of autonomous driving. In detail, the problem of autonomous intersection managements at unsignalized intersections is investigated in a complex scenario in which different classes of vehicles are involved including Connected Autonomous Vehicles, Connected Only emergency vehicles with high priority and regular unconnected vehicles. The main goal of the proposed scheme is to optimize traffic flow, ensure priorities and prevent collisions. The main contributions of the proposed approach are the novel state representation of the intersection state, which is also partially observable, the structure of the global reward function and subsequent implementation of Proximal Policy Optimization (PPO) algorithm to determine the best policy.