MARL-trained electric fish replicate real swarming behaviors

A team led by Satpreet H. Singh (Harvard, Columbia) developed a novel computational framework using multi-agent reinforcement learning (MARL) to model collectives of weakly electric fish. The agents are equipped with biophysically inspired electrosensory capabilities and actuation, trained to forage in a shared environment. The trained agents spontaneously reproduce hallmark behaviors of real electric fish, including curvilinear homing trajectories and heavy-tailed electric organ discharge (EOD) interval statistics. More strikingly, they exhibit emergent active sensing—adjusting their EOD rate based on environmental context—alongside social foraging, dominance-like asymmetries, and even aggression, all learned without explicit programming.

To understand the underlying mechanisms, the researchers performed in silico interventions such as sensor ablations, EOD silencing, and changes to food distribution. These experiments identified causal drivers of social foraging, showing how individual sensing and communication shape group dynamics. Analysis of recurrent neural network dynamics within the trained agents revealed robust encoding of task-relevant variables (e.g., food location, conspecifics’ positions) and social context. The work provides a testbed for neuroethological hypotheses in weakly electric fish and other social animals where simultaneous multi-brain recordings are infeasible, bridging AI and neuroscience with implications for swarm robotics and collective intelligence.