ANR-JCJC MIMICO:
MIMIcking the COmplexity of agent-based models
The ANR-JCJC MIMICO (2024-2027) project is funded by the french national reserch agency (ANR) under the AAPG2024 funding scheme. MIMICO is a young researcher project led by Dr. Nicolas Verstaevel, with the aims to provide next generation machine learning technics applied to agent-based modelling.
Agent-based models (ABMs) are interesting tools for modelling and studying complex phenomena in which numerous heterogeneous entities with non-linear interactions are geographically distributed and modelled at different scales. Scientists interact with ABMs by modifying the values of model parameters, either during a calibration process to produce realistic data, or to explore possible outcomes in ‘what-if’ scenarios. In-depth exploration of the parameter space of an ABM is difficult due to the relatively large number of parameters, the potentially high computational costs per model run and the non-linear relationship between parameters and results. The MIMICO project proposes to design and evaluate a new approach based on the construction of a substitution model to emulate the relationship that exists between a scenario and the results of the agent model. The novelty of the project is to design new active learning strategies, which implies that the substitution model will itself observe its learning process to request new examples. The key hypothesis is that if the learning algorithm is allowed to choose the data from which it learns, and therefore to be curious, it will perform better with less training. The surrogate model can then discover new scenarios to explore, providing interesting information for the expert while improving its own ability to mimic the ABM. The framework is being evaluated in two application areas: the first is the evaluation of the impact of new urban policies on the mobility model, and the second is the calibration of dense crowd simulation.
Interactive ABM Exploration Framework
MIMICO will introduce a novel open-source framework specifically designed for the interactive exploration of ABM models. This framework will provide researchers and practitioners with a powerful tool to investigate and understand complex ABM dynamics.
An advanced Surrogate Model Algorithm
MIMICO will develop a novel open-source surrogate model algorithm equipped with online learning capacities. This algorithm will be rigorously evaluated and compared with existing state-of-the-art algorithms, ensuring that it offers superior or comparable performance.
Advancements in Explainability
MIMICO will contribute to the field of explainability in machine learning by studying how local explainability and new metrics to evaluate self-organization can be used to design new active learning strategies, therefore enhancing our understanding of how machine learning models make decisions and adapt over time.
Real-World Applications: calibration and exploration of ABM simulations
MIMICO will demonstrate the practical utility of its contributions through applications to two real-world use cases: urban mobility and crowd simulation. Each application aims to assess the benefits of the methodology in tasks of calibration and exploration of new scenarios[NV1] . These applications will provide valuable insights into the effectiveness of the MIMICO framework and algorithm in addressing complex real-world problems.
built on Open source softwares
GAMA is an easy-to-use open source modeling and simulation environment for creating spatially explicit agent-based simulations.