From geometry to strategy: Symmetry breaking in cooperative multiagent active dynamics

We investigate the emergence of cooperative and competitive behaviors in multiagent reinforcement learning systems, where autonomous agents, modeled as active particles, learn to capture targets dynamically regenerated in geometrically distinct arenas. Using proximal policy optimization, we study circular, elliptical, and square geometries with varying target distributions. Our results show that cooperation frequently induces spontaneous symmetry breaking with agents exploring different arena regions, whereas competition tends to preserve symmetry and produce overlapping space distributions. In both fixed and mobile-target simulations, symmetry breaking occurs. The ability to break symmetry and reach optimal strategies depends sensitively on the arena geometry, agent number, and location of target-rich regions. Moreover, in a "blind" situation where agents lose direct target perception, symmetry breaking can still emerge in rare instances, highlighting that memory and interagent interactions alone may suffice to induce asymmetry. Introducing stochasticity through reward noise further facilitates this process, allowing agents to escape metastable symmetric states with moderate performance and converge toward fully asymmetric configurations with optimal performance. Together, these findings establish symmetry breaking as a key organizing principle in learning-based multiagent systems and reveal how geometry, reward structure, and stochasticity jointly govern the emergence of collective intelligence.