# SCIDYN(english)

# The SCIDYN team (Simulation, Control and Inference of agroenvironmental and biology Dynamics)

The team resulted from the evolution of the MAD team (Modelling of Agroecosystems and Decision) in September 2020.

The SCIDYN team develops and uses methods in statistics, computer science and artificial intelligence for modelling, simulating and controlling the dynamics of agricultural, biological and forestry systems.

Agricultural, biological and forestry systems are generally composed of multiple sub-systems which dynamics interact, possibly under human interventions. In order to study such systems, we develop statistical, artificial intelligence and economics modeling frameworks. We also study the algorithmic issues involved in the estimation, inference and control of these systems.

The approaches developped in the SCIDYN Team belong to the following families:

**Statistics and probabilities**. We study the probabilistic properties of stochastic models as well as the properties of statistical estimators and the development of inference algorithms. Our work in Probability theory focuses on the study of the geometry of random fields, including Gaussian fields, of eigenvalues of random matrices or of quasi-stationary states of contact processes on networks, as well as their applications in statistics. The statistical models we study are generally multi-state random processes. They can be Markovian or semi-Markovian, partially or indirectly observed (hidden models, latent-variables models, censored models...). We are interested in inference and computational statistics problems (Monte-Carlo approaches, variationnal approximation..). Theoretical problems such as identifiability or algorithmic problems concerning, e.g. dynamic network inference are also of interest to us. These theoretical advances are generally motivated by and implemented on applications in biology or ecology.

**Behavior models, decision and simulation algorithms**. The biological, agro-environmental and forest systems which we study often involve decision agents. These can be human decision-makers, exploiting the multiple services provided by the system or they can be biological entities within the system, which behavior influence the dynamics of the system (e.g. plants within a crop field). The SCIDyn Team studies behavior modelling through different approaches (Decision Theory, Sequential Decision Making, Dynamical Systems, Simulation, Reinforcement Learning...).

We develop models and algorithms for "rational" decision makers, aiming at maximizing a utility function. This can be the usual "expected utility theory", which we exploit in sequential decision under uncertainty (Markov Decision Processes) and in multi-agent decision (Non-cooperative Game Theory). We also design approaches for "qualitative" decision making (Possibility Theory). In some problems, even the use of a utility function is not adapted, in particular when simulating human behavior. Therefore, we also develop behavior models including cognitive, affective and social dimensions (Belief-Desire-Intention models, emotion models, argumentative models...). Behavior optimization is a complementary research axis to behavior modelling, which we are interested in. Indeed, rational decision models generally consist in maximizing a utility function. In the case of game theory or of "spatial" decision-making, the design of algorithms computing optimal or equilibrium behavior is a difficult problem. We tackle these problems using different algorithmic approaches: combinatorial optimization, dynamic programming, reinforcement learning... Finally, we are also interested in the integration of behavior models in the simulation of larger scope systems (agro-ecosystems, biophysical systems). We develop theoretical and algorithmic tools to perform this integration, with a simulation objective.

From an applications point of view, we tackle problems of modelling and control of biological, agro-environmental and forest systems, from the plant scale to the plot scale,farm scale or agricultural area scale. Examples of applications we are interested in are:

- [
**Plant scale**] Modelling and analysis of dynamic phenotypes at the plant or organs scale (leaves apparition process, pathogens infection dynamics, leaves motion in response to stimuli...), study of the factors and processes governing growth, reproduction or resistance to pathogens of plants.

- [
**Farm scale**] Study of the agricultural or ecological factors of crop protection and production, animal breeding or digital innovations diffusion in farming systems.

- [
**Landscape scale**] Study of the dynamics of agricultural landscapes in their biological environment: Interactions with pathogens or ecological networks, ecosystem services provision.

- [
**Forest scale**] Analysis of risk factors (storms, fires, pathogens...), preferences and multifunctionality in forest production and protection.

- [
**Ecosystem scale**] Study of the dynamics of weakly anthropised systems, with knowledge acquisition (in populations ecology) or biodiversity conservation objectives.

Publications of the SCIDyn Team can be found here: Publications

Finally, the SCIDyn Team develops and maintain some software initiated from its research work:

**Inference**

GMtoolbox: Computation of marginal probabilities in a factor graph (Matlab).

GADAG: Inference of the structure of a Directed Acyclyc Graph (R).

SISIR : variables selection in functional regression (R).

**Simulation**

VLE (Virtual Laboratory Environment): Discrete events systems simulation frameworks.

GAMA: Agent-based simulation environment.

**Decision / Optimization**

MDPtoolbox, GMDPtoolbox: Markov Decision Processes (MDP) and "spatialized" MDP solution toolboxes (Matlab).

DiceOptim, GPareto, [ GPGame] : Gaussian model functions optimization (R).

Baryonyx : Integer and Boolean Linear Programming Solver.