Proposal: Deciphering Electrocommunication with MARL and Unsupervised Machine Translation

Unsupervised machine translation (UMT) has recently been proposed as a tool for deciphering animal communication. Previous efforts, however, have attempted to align animal signals directly with human language, introducing large ecological and representational gaps that inevitably limit success. We argue that a more promising path is to generate synthetic corpora through rich, situated, biologically realistic multi-agent reinforcement learning (MARL). Such simulations yield emergent communication signals that share statistical and functional properties with real animal data, thereby narrowing the gap that hampers translation. As a case study, we present MARL agents inspired by pulse-type weakly electric fish (WEF), which rely on electric organ discharges (EODs) for both sensing and social communication. WEF provide an ideal test case because their communication signals are tightly coupled to collective behaviors such as foraging, resource sharing, and dominance interactions. Our MARL agents reproduce key features of real WEF behavior and communication, including socially aware foraging strategies, heavy-tailed EOD interval distributions, and context-dependent shifts in EOD rate. These synthetic corpora can be generated at scale, with complete access to both neural and behavioral variables, and allow for mechanistic interpretation and virtual interventions that are expensive or infeasible in vivo. We propose a methodology to combine the MARL-generated emergent communication with UMT techniques to decipher real fish EOD data. This integration opens a path toward AI-assisted deciphering of animal communication, with WEF as a proving ground and strong potential for extension to other species.