Understanding Electro-communication and Electro-sensing in Weakly Electric Fish using Multi-Agent Deep Reinforcement...

Area of Science:

• Neuroethology
• Computational Neuroscience
• Artificial Intelligence

Background:

• Weakly electric fish (e.g., Gnathonemus petersii) utilize electrical signals for sensing and communication.
• Studying their complex behaviors in natural settings is experimentally difficult.

Purpose of the Study:

• To develop a novel computational framework for studying electrosensing and electrocommunication in weakly electric fish.
• To model emergent social behaviors and neural activity using artificial agents.

Main Methods:

• Recurrent neural network (RNN) based artificial agents trained with multi-agent reinforcement learning (MARL).
• Agents modulated electric organ discharges (EODs) and movement patterns in virtual environments for collective foraging.
• Evolution-inspired rewards focused on individual fitness and emergent interactions.

Main Results:

• Trained agents exhibited emergent behaviors mirroring real fish, including heavy-tailed EOD interval distributions and context-dependent EOD shifts.
• Observed social interaction patterns like 'freeloading,' where agents reduced EOD rates while benefiting from others.
• A two-fish simulation highlighted the importance of conspecific EODs and dominance in foraging success.

Conclusions:

• The computational framework successfully simulates complex social and communicative behaviors of weakly electric fish.
• This approach provides a viable alternative for studying social animals when traditional multi-individual recordings are infeasible.
• Findings advance understanding of neuroethology and social communication in diverse animal groups.