Swarming Dynamics of Cognitive Agents

Many physical systems exhibit properties that are not found in their components. For example, biological cells form complex multicellular structures, animals display elaborate collective dynamics, and atoms compose matter. Complexity arises in very distinct scales and requires many interacting agents, making multi-agent systems (MAS)s modeling essential.
Several approaches model emergence in MAS dynamics. Some use polarized agents that self-align with their neighbors, others approaches, coarse-grain the dynamics of agents into active fluids. Game theory has also been used to model agents interacting by using predefined rules.
A limitation of these models lies in the fact that the rules are often phenomenological and do not have a clear origin. Some machine learning methods address this issue, but typically require stationary environments, which are rare in nature.
We address these issues using a modified Vicsek's model. In our model, agents are embodied with neural networks, and we apply learning via evolutionary training programming so that the agents learn to stay closer to their nearest neighbors. This is a behavior that is similar to what happens with social animals. This way, the evolutionary pressure stimulates adaptation regardless of the environmental stationarity or not.
Our results suggest that the drive to stay close to one's neighbors precedes orientation dynamics collectivity. Moreover, different migration patterns such as the formation of lanes and wave fronts are obtained by adjusting parameters like the field of vision, the turning rate, and the orientation noise, patterns that are seen in nature (lanes, lines, swarms, fronts, wave patterns, and flocking). Next steps in this research include exploring other evolutionary pressures and conditions for emergence.