Data Simulation
11 May 2023 - 12h15-14h00
Uni Battelle (Auditoire BA) & Online (Zoom)
Registration mandatory - Under this link
Program
Simulating the escape of ionizing photons from galaxies in the early Universe
Anne Verhamme, Faculty of Sciences, Department of Astronomy, Thibault Garel, Faculty of Sciences, Department of Astronomy, and RASCAS+SPHINX teams
There are two kinds of astrophysical simulations: macroscopic vs microscopic approaches, with different goals, which principles certainly have resonance in other fields. On one hand, we test theoretical concepts by implementing them at macroscopic scale to produce a statistical sample of virtual astrophysical objects -- being planetary systems, stellar populations, or galaxies – that will behave, by construction, following the theoretical expectation. And we validate/rule out the theoretical ideas by comparing statistical properties of observed and virtual populations. On the other hand, all the components in the Universe are reacting to the laws of physics that we think we understand well. So, the other approach consists in implementing these physical laws and let them act on matter and light to shape stars, planetary systems, galaxies, and the whole Universe. This is the philosophy behind this microscopic approach, but of course, there is always a minimum level of details, a scale, that we cannot resolve, and we then implement ‘macroscopic’ behaviors from theoretical expectations, to capture the effects of the physics happening at smaller scales… And the loop is closed!
I will use my work to illustrate that state-of-the-art simulations have reached a level of realism that allow us to eventually directly compare mock and real data. I argue that the most promising way to make progress with astrophysical simulations nowadays is to go all the way to produce mock observations, so that the statistical methods applied to simulations and observations are identical and results directly comparable. The transversal message that I want to emphasize is that hard work from experts was needed to make the last technical/numerical improvements: simulating the most realistic data is a probably the golden way to make progress on many science problems, but achieving this goal certainly implies multi-disciplinary collaboration.
Simulating Cooperation & Resilience with Agent-Based Models
Thomas Maillart, Geneva School of Economics and Management, Information Science Institute
In nature and society, adaptation and resilience are key to the survival of species. While adaptation has long been considered as an evolutionary matter, there is now much evidence that individuals transfer genes horizontally (e.g., epigenetics). Further, humans strive as social beings, by deriving a competitive advantage from social interactions and the collective integration of knowledge that ensues. New institutionalism economists have largely documented how humans manage to organize collective action in order to overcome challenging problems that could only be solved as the whole is more than the sum of individual contributions. Here, we posit that the more unpredictable, time-critical, sharply evolving and/or existential challenges faced, the more humans need to cooperate to overcome and adapt. Such challenges include cybersecurity risks, the consequences of climate change and the degradation of biodiversity. While some empirical evidence exists of the emergence of cooperation in populations of monkeys in the face of a disaster, there is a dearth of knowledge regarding the fundamental mechanisms of resilience through cooperation. To investigate the mechanisms of cooperation emergence in the face of dynamic threat landscapes, we have developed an agent-based model, which accounts for the intricate relation between cooperation and resilience. Here, the use of simulation is justified by the absence of hard empirical evidence on the one hand, and on the other hand, by the difficulty of eliciting a closed form mathematical theory, largely because of endogeneity and self-organization.
Simulating domain-credible data
Stéphane Marchand-Maillet, Faculty of Sciences, Departement of Computer Science
Training and benchmarking data science processes require data. In domains such as medicine where data is scarce and either difficult or expensive to obtain, generating data artificially is required. However, in domains where the data model is constrained, rough randomly generated data is not credible. This is again the case in medical data analysis where data should abide biological processes and constraints. Data simulation should therefore build itself over real data modeling even if scarce. We briefly present the case of Flow Cytometry data analysis where generating credible data would allow for the design of a parameterized benchmark.
