Most natural collective behaviors allow a group to accomplish complex tasks that a single individual could not achieve¹ (coordination, long-distance visibility, intimidation of predators, etc.). The study by physicists of emergent patterns related to collective behavior - even if it is essential - has somewhat obscured this primordial function while these “intelligent” behaviors find various applications, in particular in swarm robotics². Indeed hundreds of simple and inexpensive devices can be deployed to perform tasks that a single robot, even very sophisticated, could not accomplish; this is the case, for example, of the search for a target in a complex environment.

By studying the collective behavior of agents controlled by simple neural networks, it has been observed that they are able to collectively find the solution of mazes, even though the only information shared between the agents is to estimate the density of other agents:

However, this behavior is not generalizable to all possible mazes, which probably stems from the fact that for these simulations the weights of the control neural network were adjusted by hand. 

In this internship it is proposed to evolve neural networks controlling a population of agents in order to automatically find architectures capable of solving any maze, and thus demonstrate that this intelligent behavior can emerge even when the communication between agents is reduced to its minimum. For this, evolutionary algorithms very close to natural evolution will be used^3,4.

Bibliography

1.Ioannou, C. C. & Laskowski, K. L. A multi-scale review of the dynamics of collective behaviour: from rapid responses to ontogeny and evolution. Philos. Trans. R. Soc. Lond. B Biol. Sci.378, 20220059 (2023).

2.Schranz, M., Umlauft, M., Sende, M. & Elmenreich, W. Swarm Robotic Behaviors and Current Applications. Front Robot AI7, 36 (2020).

3.Stanley, K. O. & Miikkulainen, R. Evolving neural networks through augmenting topologies. Evol. Comput.10, 99–127 (2002).

4.Mattiussi, C. & Floreano, D. Analog genetic encoding for the evolution of circuits and networks. IEEE Trans. Evol. Comput.11, 596–607 (2007).