Activities, influence and research work

Activities of Research / Research work by Guillaume GUERARD. Projects, seminars, conferences, conferences, internship supervision, all the latest news on research activities.


View Scholar profile

View Publons profile

Research activity

Thesis supervision

  • Loup-Noé LEVY: 2020-2023 CIFRE Energisme
  • Ihab TALEB: 2021-2024 H2020 MAESHA


  • 2021-2024: Loc TRAN, EPHE – Hypergraph clustering method
  • 2019-2022: Laureen DEMAN, Rhones-Alpes – Economic optimization of the use of hydroelectric dams
  • 2018-2021: Tuan TRAN, EPHE – Bayesian approach and learning for the modeling of a class of economic problems and their application in finance
  • 2017-2020: Mahdjouba AKERMA, INREA – White box and black box methods for predicting energy consumption

Seminars and Conferences

Master Thesis supervision

2022: Louis LESIAK and Pacôme MANCEAUX and Aurélien AURUS, various work on Energy 4 CLimate

2022: Emma KARSENTI, detection of polymorphic malware by dynamic analysis

2022: Clément CORNET and Maxence CHOUFA, clustering algorithm by pretopology

2022: Paco BARRA, Systemic modeling of the island of Mayotte

2021: Benjamin SLIOSBERG, Method predictive of human impacts of natural disasters

2021: Rémy SIMINEL, Human gene embedding using global vectors algorithm

2021: Flavien DESEURE-CHARRON, Behavioral analysis of photographic tourism

2020: Jeremy DONADIO, machine learning approach for polymorphic virus activity detection.

2020: Quentin GABOT, data mining in a Big Data framework, application to tourism

2020: Antoine THIOL, hybrid deep learning method for time series prediction

2020: Lucas BICHET, security vulnerability detection algorithm in web browsers

2019: Loïc STEINLE, statistical analysis of Instagram tourist hotspots

2019: Théo DEMESSANCE, Markovian prediction of tourist flows

2019: Maelys TESTA and Floriane GIRAULT, quantification and qualification of energy comfort

2019: Thomas REYNOLDS, clustering and analysis of tourist travel in Paris

2018: Mathilde LAVELLE, analysis of multi-agent smart grid models

2018: Lilia BEN BACCAR, analysis of tourist behavior using data mining

2018: Victor SIROT, analysis of tourist behavior using Markovian process

2018: Victoire BONNET, feasibility study of a blockchain for heat networks

2017: Loup-Noé LEVY and Bastien PICHON, digital modeling of home automation

2017: Rémi TOLLERET, usefulness of BIM for energy efficiency

2017: Hugo POUSSEUR, multi-agent model of an eco-district under JADE

2016: Zeinab NEHAI, the blockchain for self-consumption

2016: Loup-Noé LEVY, multi-agent model of an eco-district

2016: Estelle CATHELINEAU, diffusion of energy in a network

Master’s/internship supervision

  • 2022-2023

Guillaume GALLET: smishing detection tools

  • 2021-2022

Benjamin SLOSBERG: multi-agent modeling of human contentment in temporary work

  • 2020-2021

Quentin GABOT: profiling tourist behavior

Sumayya CHABANE (with Supergrid-Institute): optimization by multi-agents of a hydraulic power plant with energy market

  • 2018-2019

Lilia BEN BACCAR: pattern mining for predicting tourist behavior

  • 2017-2018:

Hugo POUSSEUR: multi-agent model of an eco-district

Manon RIVOIRE: modeling of home automation through learning

Marc RUAULT: game networks applied to an eco-neighborhood

  • 2015-2016:

Zeinab NEHAI: request-response and home automation

Bastien PICHON: demand-side management and demand-response

Teaching activities

Scientific theme: modeling of complex systems


Since the 1950s, it has become clear to scientists that a new way of thinking about and analyzing a system is essential to understanding many of the complex challenges facing humanity. The new paradigm involves inferring rules about how the dynamic behavior of a complex system depends on the combined properties of individual elements, the nature of interactions between elements, as well as the topology of interactions between elements, in order to understand and predict these systems and control them to have desirable properties.

For example, it is important to know which characteristics of complex systems generate resilience against disturbances versus properties that improve the sensitivity of the system and allow it to move to a different equilibrium state for a wide range of questions, for example on climate change and the collapse of ecosystems, the evolution of species and microorganisms, energy systems or even human displacement/migration.

Complex systems are studied from four complementary angles: emergence, resilience, phase transitions and predictability/control. These abstract unifying concepts are linked across various disciplines; mathematical models using numerous fields of research as presented in Figure 1.1.

système complexe

Before talking about the two complex systems studied (electricity network and tourist flow), this introduction will allow readers to better familiarize themselves with the different concepts related to complex systems.

Formal system and empirical system

It is not possible to talk about complex systems without talking about Shannon entropy and chaos theory. The first is a mathematical function which, intuitively, corresponds to the amount of information contained or delivered by an information source. The entropy of a complex system has a particular behavior. Indeed, the entropy of the system as a whole is greater than the sum of the entropies of each of its subsystems. From a mathematical point of view, the overall system has more degrees of freedom (emergent properties) than all of its subsystems.

A system falls under chaos theory if it is extremely sensitive to small causes and if its behavior has a cyclical aspect. For example, weather systems are very sensitive to small causes and their dynamics are cyclical: the seasons. Each behavior of the system is a point defined in “n”-dimensional spaces, called phase space.

Trajectories in phase space converge towards areas called attractors. The same system can have one or more attractors. Attractors are important because they represent stable zones in the sense that the system tends to join these zones or to migrate to another stable zone. A basin of attraction is a space of points which tend to join irremediably towards an attractor.

The systems that fall under chaos theory come in two forms: formal systems and empirical systems. The former are defined mathematically, often in the form of nonlinear differential equations, for example the Lorenz equations.

Empirical systems are unpredictable because of their sensitivity to initial conditions that it is impossible to rigorously define, or even count. These systems evolve towards basins of attraction, but their dynamics within these basins are unpredictable.

Empirical systems are observational and cannot be modeled mathematically in an accurate manner. To better understand this notion, they are often compared to an elephant that blind people try to recognize by touch (figure 1.2); note that in this example the global system is known, which is rarely the case.

système complexe

The study of complex systems has been concretized and formalized by the work of the Santa Fe Institute since 1984. This institute's mission is to research laws common to complex systems of varied natures; to define analysis and forecasting tools. A complex system is defined as follows:

A complex system is made up of agents that interact with each other, with their environment and with the emergent phenomena created by these interactions. The agents can be of varied nature: an animal, a person, a group of people, an institution, an organ, a cell, an enzyme. An agent's behavioral rules define the stimuli it emits to other agents based on the stimuli it receives from other agents and its environment. These rules evolve according to the agent's experience: stimuli that it has received and that it has emitted. Emergence is a process of creating phenomena through the interactions of agents: among themselves and with their environment. Emerging phenomena are of varied nature, for example the appearance of a new agent, a modification of the environment, a law of distribution of events. Emergences are not planned or managed by an authority that would have an overall view of the system.

Complex systems can be of different types. My work focuses on artificial systems (artifacts). These are systems built by humans mainly using computers, such as: electricity distribution networks, fuloscopy, swarm robotics. These systems have the potential for functionalities of complex adaptive systems.

Characteristics of complex systems

Table 1.1 compares the Cartesian approach with a systems approach. The different characteristics of the complex systems [20] present in figure 1.3 will be detailed later.

système complexe

systèmes complexes

Complex systems work from the bottom up. It is the agents at the bottom of the hierarchy who make the systems work, who produce emergent phenomena through interactions between them and with their environment.

The role of the top of the hierarchy is limited to creating conditions favorable to the emergence of the desired phenomena (basin of attraction). The capacities for adaptation and innovation of an organization thus decentralized are notably greater than those of a centralized structure.

Complex systems cannot be divided, remember that entropy tells us that “the whole is more than the sum of the parts”. To study a complex system, all of its components must be considered simultaneously: agents, interactions and the environment. It is also necessary to take into account the interactions and the temporal evolution of the system.

It is necessary to bring together multidisciplinary skills that cover all facets of the system studied. This arises from holism which does not allow properties linked to different disciplines to be studied in isolation. As a reminder, Figure 1.1 presents various theories related to all complex systems.

Agents have different properties, laws, rules, behaviors and actions. Agents can however be grouped by class. Very often the specification of agents is of the order of UML diagrams in 4+1 views. The diversity of agents reinforces the properties of complex adaptive systems: emergence, innovation, self-organization.

Medium and long-term developments in complex systems are unpredictable because it is impossible to define all the variables with the precision required for a forecast. Accuracy requirements grow exponentially with the scope of forecasts.

It can be futile to try to identify the cause of a situation observed in a system because this situation is often due to multiple causes. It is impossible to identify a main cause among these multiple causes which have been the subject of a succession of amplifications and attenuations due to feedbacks.

For example, the causes of the rise or weakening of a civilization are often uncertain with controversies fueled by numerous theses. It is necessary to distinguish the accidental cause which triggered an event from all the causes which created the latent state which allowed this triggering. This accidental cause is easily identifiable and is sometimes wrongly considered as the main cause of this event.

Complex systems are subject to shifts which are sudden developments in terms of their scale and speed. For example, a collective enthusiasm, an economic crisis, a revolution, or the initial conditions. A tipping point is a state of a system where a small cause can cause a profound and abrupt change in the state of the system.

Like any chaotic system, a complex system has cycles of evolution between the different basins of attractions. A shift can be conceptualized as a jump from one basin of attraction to another. We often do not know how to identify the attractors of complex systems.

Feedback occurs when an agent receives stimuli that are influenced by the stimuli it emitted. Feedbacks generate the fundamental properties of complex systems: holism, convergence towards basins of attraction, cyclical dynamics, bifurcations, shifts, evolution, adaptation, emergence, self-organization, etc. Feedbacks can have a stabilizing effect, for example regulating supply by demand in a market, or on the contrary an amplifying effect, for example creating booms whose archetype is a speculative bubble. The adaptive capacities of complex systems are due to feedback which causes the rules of behavior of their agents to evolve based on their experiences.

Combinations of these characteristics lead to the main characteristic of all complex systems: emergence. Emergence is the creation (of a holistic view) or self-organization (of a micro view) of new characters or phenomena in the system. Emergence is necessary for adaptation, evolution, coevolution, reproduction.

This occurs in various forms:

1. an adaptation/evolution of the characteristics of agents in space and time;
2. an adaptation/evolution of behavior and rules of agents in space and time;
3. self-organization with specialization of agents;
4. self-organization with hierarchy of agents;
5. coevolution between a pair or more agents;
6. a creation of new agents by evolution and specialization of existing agents.

Systemic study and modeling

The study of complex systems has different objectives:

• Determine the agents;
• Understand their functioning, their laws;
• Understand their evolution;
• Predict their developments;
• Define interventions to make them evolve in the desired direction.

Systemic study has a wide range of methods and tools:

1. agent-based simulation
2. graph theory and networks, collective intelligence, analogy with other complex systems
3. chaos edge theory, morphogenesis, chaos theory, catastrophe theory, fuzzy variable logic, memetics
4. operational research and metaheuristic algorithms (see figure 1.4).

aide à la décision

These tools work iteratively. They calculate the state of a system at time ti+1 based on its entire state at the previous time ti. They take into account the entire state of the system considered: states of the agents and the environment, stimuli emitted by the agents and the environment.

Agent-based simulation is a response to the excessive complexity of solving through mathematics in the study of complex systems [10]. It is often even impossible to put them into an equation because of the variety of agents and their rules of behavior, the evolution of the rules of behavior of agents according to their experiences. At any time all agents can have different rules of behavior. Complete modeling of cognitive agents, people or groups of people, is impossible. We must therefore simplify while retaining what is relevant for the phenomena studied. The modeling of complex systems are decision support tools.

The process which leads to deciding on an action implements a set of convergent operations, logical or not, on a more or less important and relevant group of information, based on a set of knowledge, in an environment determined in order to obtain a result. The relevance of the procedure followed to make a decision is rarely evaluated because it is very complex, only the result is in relation to an initially sought objective.

Decision aids are therefore operations that facilitate the decision-making task by simplifying or shortening the cognitive and mathematical path followed by agents. The functions of these aids can be very diverse [23]:

• searches for relevant information;
• organization of information;
• partial or total processing of disjoint sets of information;
• ordered activation of knowledge;
• establishment of scenarios;
• spatial and/or temporal representations;
• proposed decisions.

Agents, cognitive or reactive, all use decision support processes in order to adopt a particular behavior (see Figure 1.5) [26]. All of these decisions cause, disrupt or vary the emergence, resilience, phases and control of the entire system.

agent cognitif

The modeling of a complex system can vary depending on the decision sought. A generic model then leads to a decision being made on all aspects of this system.


Interdisciplinary management in complex systems


Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Interdisciplinary project management


MLOps methodologies applied to complex systems

LNL and digital twin

Application to Smart Grid


Energy development has brought about a paradigm shift in the 21st century. Industrial and political entities and the scientific community seek to make cities and the network smarter \cite{ramchurn2012putting}. The French Institute of Demographic Studies (INED) states that the world population will increase to 10 billion in 2050, while the urban population will double (an increase of 63\%). Expanded urbanization requires new ways of understanding and managing the complexity of energy production and consumption.

Today, although cities occupy only 2\% of the planet's surface, they are home to approximately 50\% of the world's population, consume 75\% of the total energy generated and are responsible for 80\% of the greenhouse effect \footnote{United Nations Environment Program, Visions for Change. Recommendations for effective policies on sustainable living, 2011}. The development of smart cities depends on the level of intelligence of electricity networks, from producers to consumers, from consumers to producers. The most important aspect is coordination between all entities; a micro-grid will be able to encourage consumers to modify their consumption under critical conditions so as not to alter the electrical infrastructure.

Around the world, smart grids are being developed to reduce electrical waste and prevent power outages. Simulating a microgrid, an eco-district or a virtual power plant is difficult, given their different behaviors and structures. Each varies according to several aspects: social, economic, energy, mobility and well-being of its inhabitants.

Smart grids and smart cities need to be understood. They are generally described as a complex system and the best way to analyze it is through modeling and simulation. In this context, the objective is to create a context-free model in order to provide decision support tools for social, economic and algorithmic questions on a smarter microgrid.

Systemic modeling

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Decision-making tools

The network constantly needs to estimate its consumption / production in the near and distant future. Formerly based on data from previous days and years, the smart grid must be equipped with prediction systems that analyze data in real time.

For this, many machine learning tools, and more particularly deep learning, are based on data from hundreds of sensors in order to make a reliable prediction. However, this system is complex in its implementation, operation and maintenance.

The objective of the work carried out on this subject is to provide a reliable consumption prediction tool without having to resort to conditioning a building. The data used can be provided by a simple Linky type smart box.

The consumption of each device is transformed into consumption sequences. The latter then provide sufficient data for the prediction. The models used so far come from data mining and are based on learning by prefix trees. The models are Markovian types (with or without grammatical inference) or compact trees.

Various applications

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Application to Tourism


Due to the rapid exchange of information on the Internet by users, many destinations have experienced an increase in the number of tourists. Businesses recognized Instagram's enormous potential as an advertising platform early on, with many savvy businesses collaborating with the site's top influencers to serve as ambassadors \cite{fatanti2015beyond}. New Zealand's Lake Wanaka Tourist Board is a good example: in 2015, they invited Instagram influencers to share their photos of the area. As a result, tourist visits jumped by 14\%, representing New Zealand's fastest growth rate.

Analyzing tourist sites is quite easy, but analyzing tourist flows between hotspots in the same city is challenging \cite{li2018big}. The capacity of our current processing technologies and algorithms is blocked by three aspects: data volume, data generation speed, variety of data types, also known as 3V.

Tourists want a recommendation based on their interests because time is a key part of travel planning and it is a time-consuming task when carried out by the tourist.

Tourist segmentation

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Analysis of movements

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Common tourist interests

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.

Application in Cybersecurity


The network's overload of digital tools makes its entire system non-resilient to the threat of computer viruses. The number of attacks by polymorphic viruses or BOTNET increases as the system migrates from physical systems to digital systems.

The detection of polymorphic viruses is complex or even impossible without knowledge of the network and its behavior. Detection of suspicious activity is then possible by comparing reality with the system's different predictions.