Inverse MAS Design

Most Multi-Agent Systems (MAS) utilize agents to simulate complex environments, with the primary goal of observing global behaviors that emerge from local interactions — much of this research falls under the umbrellas of Artificial Life or Collective Robotics. Alternatively, MAS scholars tweak and optimize agent parameters to achieve a desired global behavior — this is usually framed under Swarm Intelligence.

I’m trying to do something quite different.

• Starting with a known number of agents and a specific target global behavior, the goal is to evolve the agent architectures and external stimuli capable of producing the behavior. Similarly to Occam’s Razor: the simpler, the better.

• Then, we will compare these synthetic stimuli with the real-world data gathered by my colleagues, expert in animal behavior, and we will possibly refine the agent structures.

• Eventually, we shall try to map the internal structures of the synthetic agents with the physical attributes that could be measured in the real, biological subjects, such as the dopamine levels or other characteristics.

This research is conducted in collaboration with: AgroParisTech (Paris-Saclay University), the French National Centre for Scientific Research, the Nencki Institute of Experimental Biology, and University of Cagliari. We are currently applying for an EU grant; should the funding be secured, the project will scale up into a multi-year initiative.

I am seeking motivated Master’s students in Computer Science or Data Science – as the generation of agents will rely on Evolutionary Computation, students from my Computational Intelligence are ideal candidates. Anyhow, if you feel interested, please contact me.