Short bio

Dr. Mihail Mihaylov is currently a senior research engineer at Enervalis, Belgium working in the domain of smart grids and blockchain technology. Previously he was a post-doctoral researcher consecutively in three institutions: Vrije Universiteit Brussel, Belgium (2 years), Sensing & Control Systems, Spain (2 years) and University of Leuven, Belgium (1 year). In 2012 he obtained his Ph.D. degree in artificial intelligence with highest distinction from Vrije Universiteit Brussel, Belgium with his doctoral thesis "Decentralized Coordination in Multi-agent Systems". In 2008 he obtained his M.Sc. degree in artificial intelligence with distinction from Maastricht University, The Netherlands and subsequently was awarded Leo Coolen Award for his thesis entitled "Computational Mechanism Design for Wireless Sensor Networks". He received his B.Sc. in computer engineering from University of Duisburg-Essen, Germany in 2007. Presently, his main interests and work focus on renewable energy, smart grids and blockchain technology.

Ph.D. in Artificial Intelligence

2008-2012, Vrije Universiteit Brussel, Belgium

Thesis title: "Decentralized Coordination in Multi-agent Systems". Graduated with Highest Distinction.

Focus: machine learning, wireless sensor networks, game theory, decentralized algorithms, multi-agent systems, collective behavior.

  • Abstract - click to see abstract

    Many computer systems are comprised of multiple entities (or agents) with common objectives. Though these systems can be made intelligent, using artificial intelligence techniques, individual agents are often restricted in their capabilities and have only limited knowledge of their environment. However, the group as a whole is capable of executing more complex tasks than a single agent can perform. Individual agents, therefore, need to coordinate their activities in order to meet the design objectives of the entire system. Implementing a centralized control for distributed computer systems is an expensive task due to the high computational costs, the communication overhead, the curse of dimensionality and the single point of failure problem. The complexity of centralized control can be reduced by addressing the problem from a multi-agent perspective. Moreover, many real-world problems are inherently decentralized, where individual agents are simply unable to fulfill their design objectives on their own. In multi-agent systems with no central control, agents need to efficiently coordinate their behavior in a decentralized and self-organizing way in order to achieve their common, but complex design objectives. Therefore it is the task of the system designer to implement efficient mechanisms that enable the decentralized coordination between highly constrained agents.

    Our research on decentralized coordination is inspired by the challenging domain of wireless sensor networks (WSNs). The WSN problem requires resource-constrained sensor nodes to coordinate their actions, in order to improve message throughput, and at the same time to anti-coordinate, in order to reduce communication interference. Throughout this thesis we analyze this (anti-)coordination problem by studying its two building blocks separately so that we form a solid basis for understanding the more complex task of (anti-)coordination. We study pure coordination in the problem of convention emergence and pure anti-coordination in dispersion games. We then study the full problem of (anti-)coordination in time, as seen in the WSN domain.

    Our main contribution is to propose a simple decentralized reinforcement learning approach, called Win-Stay Lose-probabilistic-Shift (WSLpS), that allows highly constrained agents to efficiently coordinate their behavior imposing minimal system requirements and overhead. We demonstrate that global coordination can emerge from simple and local interactions without the need of central control or any form of explicit coordination. Despite its simplicity, WSLpS quickly achieves efficient collective behavior both in pure coordination games and in pure anti-coordination games. We use our approach in the design of an adaptive low-cost communication protocol, called DESYDE, which achieves efficient wake-up scheduling in wireless sensor networks. In this way we demonstrate how a simple and versatile approach achieves efficient decentralized coordination in real-world multi-agent systems.

M.Sc. in Artificial Intelligence

2007-2008, Maastricht University, The Netherlands

Thesis title: "Computational Mechanism Design for Wireless Sensor Networks". Received Leo Coolen Award for thesis. Graduated with distinction.

Focus: mechanism design, wireless sensor networks, game theory, data mining, machine learning, probabilistic robotics, intelligent search techniques, optimization algorithms.

  • Abstract - click to see abstract

    In this thesis we provide a general overview of Computational Mechanism Design and study its practical application to the energy conservation problem in Wireless Sensor Networks. Mechanism Design is an approach to Game Theory that studies how to make agents achieve the designer's goal out of their own self-interest. Our approach to this problem is based on the Collective Intelligence framework of Wolpert et al, which we regard as Learnable Mechanism Design. COllective INtelligence (COIN) describes how selfish agents can learn to optimize their own performance, so that the performance of the global system is increased.

    In our research we study the application of the COIN framework to Wireless Sensor Networks (WSNs), which are collections of densely deployed sensor nodes that gather environmental data. The main challenges in WSN design are the limited power supply of nodes and the need for decentralized control. Therefore, our aim is to increase the autonomous lifetime of the network in a decentralized manner, by making each sensor node use a learning function to optimize its own energy efficiency, so that the energy efficiency of the global system is increased. We show that nodes in WSNs are able to develop an energy saving behaviour on their own, when using the COIN framework. We study the performance of different learning algorithms in a simulation environment and provide guidelines on the choice of algorithm for energy efficiency optimization in WSNs for different domains.

B.Sc. in Computer Engineering

2003-2007, Universität Duisburg-Essen, Germany

Thesis title: "Mathematical Methods to Ascertain Flow-Curve Functions With the Help of Neural Networks".

Focus: Software development and testing, usability engineering, logical devices and digital systems, neural networks, regression analysis, basic electronic devices, design theory, physics.

  • Abstract - click to see abstract

    Computations of cold and hot flow curves have long been used in metal forming via isolated mathematical functions, that make the calculation of smooth and precise material flow curves extremely difficult for wide temperature ranges. When modern statistical methods are integrated into a computer environment in the form of software application, accurate flow curves for different steel grades can be approximated. These methods, however, have their advantages and disadvantages, but when combined in a proper way, certain model arises, which puts together the benefits and eliminates most of the drawbacks of these techniques.

    This document explores the application of regression analysis and artificial neural networks in metal forming in order one to obtain values of flow stress for the range 20° – 1200° C. The computation of flow stress is carried out by software – “Flow Stress Plotter”, developed by the author, having also the purpose to visualize flow curves in two as well as in three dimensions.