PhD thesis offer: Combining co-simulation with optimization and learning.

Combining co-simulation with optimization and learning.


Laboratory: LORIA Team: SIMBIOT Department 3

Supervisor:         Vincent CHEVRIER           vincent DOT chevrier AT



MECSYCO (Multi-agent Environment for Complex System CO-simulation) is a middleware for co-simulation [Camus et al 18, Paris 19]. It allows several models and their simulators to interact to simulate a global system designed from different components. Each entity or phenomenon of the target system is modelled via a dedicated software, then the different models interact to account for the behaviour of the system as a whole, we speak then of multi-model co-simulation.

The interest of co-simulation [Gomes et al 18]  (rather than simulation) is to be able to integrate heterogeneous models (different simulators, different formalisms, …) in an incremental process to study the global system. Co-simulation allows for example to study an electrical distribution system as composed of an electrical part, a telecommunication part and a control part by jointly simulating these parts.

Mecsyco is used for different studies whose central theme is energy: : for example, the Territory, Industry and Energy Chair of the University of Lorraine in partnership with EPFL (processes energy and energy synergies); the APHEEN-LORIA collaboration (energy renovation of existing building), the ULHyS (LUE) project (study of hydrogen microgrids) in collaboration with the GREEN laboratory, etc. See for a non-exhaustive list of other examples of application.

The use of co-simulation is generally part of a larger decision support approach: a system is studied to make a decision, such as sizing (optimizing) equipments (e.g. researching the specifications of an energy storage) according to constraints (local renewable energy production, weather and thermal behavior of a house, uses and usersetc.). This is, for example, one of the scientific objectives of the Territory, Industry and Energy Chair.

Co-simulation also assumes the existence of models (e.g. thermal model of the house).  However, it happens that for some systems the model of a component does not exist but data are available. . In this case, the model of this component could be built from the data by learning methods. This is for example one of the objectives of the work undertaken with APHEEN in [Vernerey 20] in which a model of the thermal behaviour of a flat was built from weather data and temperature readings in the flat.

Subject of the thesis

The thesis aims to explore the relationships between co-simulation, optimization and learning having in mind to propose a global decision support approach based on co-simulation.

The relationships between co-simulation, optimisation and learning can be considered in multiple ways.  A first point of view can be to consider optimisation and learning as “beside” the co-simulation tool and being able to execute it to obtain data and/or to modify the parameters or the structure of the multi-model.  These relationships can also be conceived at different levels: at the level of a component/model (e.g. optimising energy storage, learning the thermal model of the flat) or at the level of the composite/system (e.g. learning the behaviour of a system in order to have a simplified model that is quicker to simulate (model reduction) or simpler to manipulate; optimising the behaviour of the system). Another point of view is to consider that optimisation and learning are part of the co-simulation as a component (and therefore as a model) of decision within the simulated system (e.g. the model of an energy management system).

As presented above, the notions of optimisation and learning are very different in nature (particularly in terms of the problems to be solved and the techniques to be used) depending on the point of view adopted. This is why a first stage of the work will be to clarify these notions and to put them in relation to a case study.


The central idea of study is to question the DEVS formalism and the associated concept of System, Entity Structure in order to assess to what extent they are compatible with these different notions of optimization and learning, and under what assumptions according to each of the facets compatibility can exist.

A first (restricted) case study will be chosen and confronted with a first proposal of definitions, then in an incremental way, the proposal will be improved  and again confronted with one or several case studies.

The MECSYCO platform ( will be used as a software base for the integration of the proposals.

Expected profile

The candidate must have a master degree in computer science (or equivalent) with research experience. Skills in one of the areas of (co-)simulation, learning or optimization will be an advantage.

Both conceptual and software skills are valuable as the thesis goes from TMS/DEVS concepts to their implementation in MECSYCO.


Introductive bibliography

Paris Thomas ” Modelling of complex systems by composition ” PhD thesis from the University of Lorraine 2019. (in French).

Gomes al ” Cosimulation: a survey, ACM Computing Surveys May 2018 Article No.: 49

Camus et al Co-simulation of cyber-physical systems using a DEVS wrapping strategy in the MECSYCO middleware SIMULATION, SAGE Publications, 2018, 94 (12), pp.1099-1127.

Thomas Paris, et al. “Teaching cosimulation basics through practice”. In: 2019 SUMMER SIMULATION CONFERENCE. Berlin, Germany, July 2019.

A description of the MECSYCO tool is available on, as well as more complete references on the work undertaken.


Logo d'Inria