Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Satpreet H. Singh, Sonja Johnson-Yu, Zhouyang Lu, Aaron Walsman, Federico Pedraja, Denis Turcu, Pratyusha Sharma, Naomi Saphra, Nathaniel B. Sawtell, Kanaka Rajan

TL;DR

The paper tackles the challenge of understanding electro-communication and electrosensing in weakly electric fish by introducing a biologically inspired multi-agent reinforcement learning framework. It trains recurrent neural network agents to emit and sense electric organ discharges in a simulated foraging arena, reproducing key social and sensing phenomena such as heavy-tailed EOD intervals and freeloading, and uses a minimal two-fish assay to isolate electro-communication. Key contributions include a detailed MARL implementation with shared RNNs, PPO training, biomimetic sensing channels, and interpretable internal dynamics, enabling targeted ablations and steering experiments. The work provides a versatile, in silico platform to generate hypotheses about electric fish communication and to produce synthetic datasets for neuroethology and related fields, potentially guiding future experimental studies while reducing invasive data collection.

Abstract

Weakly electric fish, like Gnathonemus petersii, use a remarkable electrical modality for active sensing and communication, but studying their rich electrosensing and electrocommunication behavior and associated neural activity in naturalistic settings remains experimentally challenging. Here, we present a novel biologically-inspired computational framework to study these behaviors, where recurrent neural network (RNN) based artificial agents trained via multi-agent reinforcement learning (MARL) learn to modulate their electric organ discharges (EODs) and movement patterns to collectively forage in virtual environments. Trained agents demonstrate several emergent features consistent with real fish collectives, including heavy tailed EOD interval distributions, environmental context dependent shifts in EOD interval distributions, and social interaction patterns like freeloading, where agents reduce their EOD rates while benefiting from neighboring agents' active sensing. A minimal two-fish assay further isolates the role of electro-communication, showing that access to conspecific EODs and relative dominance jointly shape foraging success. Notably, these behaviors emerge through evolution-inspired rewards for individual fitness and emergent inter-agent interactions, rather than through rewarding agents explicitly for social interactions. Our work has broad implications for the neuroethology of weakly electric fish, as well as other social, communicating animals in which extensive recordings from multiple individuals, and thus traditional data-driven modeling, are infeasible.

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

TL;DR

The paper tackles the challenge of understanding electro-communication and electrosensing in weakly electric fish by introducing a biologically inspired multi-agent reinforcement learning framework. It trains recurrent neural network agents to emit and sense electric organ discharges in a simulated foraging arena, reproducing key social and sensing phenomena such as heavy-tailed EOD intervals and freeloading, and uses a minimal two-fish assay to isolate electro-communication. Key contributions include a detailed MARL implementation with shared RNNs, PPO training, biomimetic sensing channels, and interpretable internal dynamics, enabling targeted ablations and steering experiments. The work provides a versatile, in silico platform to generate hypotheses about electric fish communication and to produce synthetic datasets for neuroethology and related fields, potentially guiding future experimental studies while reducing invasive data collection.

Abstract

Weakly electric fish, like Gnathonemus petersii, use a remarkable electrical modality for active sensing and communication, but studying their rich electrosensing and electrocommunication behavior and associated neural activity in naturalistic settings remains experimentally challenging. Here, we present a novel biologically-inspired computational framework to study these behaviors, where recurrent neural network (RNN) based artificial agents trained via multi-agent reinforcement learning (MARL) learn to modulate their electric organ discharges (EODs) and movement patterns to collectively forage in virtual environments. Trained agents demonstrate several emergent features consistent with real fish collectives, including heavy tailed EOD interval distributions, environmental context dependent shifts in EOD interval distributions, and social interaction patterns like freeloading, where agents reduce their EOD rates while benefiting from neighboring agents' active sensing. A minimal two-fish assay further isolates the role of electro-communication, showing that access to conspecific EODs and relative dominance jointly shape foraging success. Notably, these behaviors emerge through evolution-inspired rewards for individual fitness and emergent inter-agent interactions, rather than through rewarding agents explicitly for social interactions. Our work has broad implications for the neuroethology of weakly electric fish, as well as other social, communicating animals in which extensive recordings from multiple individuals, and thus traditional data-driven modeling, are infeasible.

Paper Structure

This paper contains 4 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Analyses
  4. Discussion

Figures (3)

  • Figure 1: Overview of our MARL framework for modeling weakly electric fish communication. (a) Schematic of the training loop, where agents interact with a simulated arena, emitting and sensing electric organ discharges (EODs) through weakly electric fish-inspired sensors. Rewards encourage successful foraging and penalize aggressive encounters. (b) Example trajectories from four agents in a single foraging episode, showing exploration and food acquisition. (c) Snapshot of the arena (top) showing agents, food sources, and simulated electric fields; bottom shows temporally-structured EOD spike trains across individual agents. (d) Sequential Pulse Interval (SPI) distributions from real fish (left) and MARL-trained agents (right), showing that in silico agents reproduce the heavy-tailed statistics observed in biological data. Insets show log-linear curves compared to empirical curve fits.
  • Figure 2: (a) Comparison of EOD probabilities under various conditions: (Left to Right) (a1) Effect of the Knollenorgan in competitive environments. The presence of the Knollenorgan (which provides long-range information about other agents) increases EOD rates in competitive scenarios only, suggesting the importance of social information in limited-resource regimes. (a2) Effect of the Knollenorgan in non-competitive environments. The Knollenorgan has no impact on EOD rates in non-competitive scenarios, suggesting that long-range information is not important when food is abundant. (a3) Agents (with collective sensing and Knollenorgan-enabled long-range sensing) generally tend to produce more EODs in arenas with limited food ("Competition") compared to cases where food supply is unlimited ("No competition"), suggesting that competition drives higher EOD rates as agents actively search for food. (a4) Effect of collective sensing in non-competitive environments, aggregated with and without Knollenorgan-based long-range sensing. Collective sensing reduces EOD rates as agents can gather short-range information from the EOD discharges of their neighbors, consistent with pedrajaCollectiveSensingElectric2024. (b) Agent displacement over a $\approx 0.36$ second window in non-competitive environments indicates that long-range social information associated with larger movement bouts, potentially facilitating more extensive spatial exploration and thereby more efficient foraging. (c) Inter-agent inequality in food consumption increases under food scarcity, as measured by Theil Index conceiccao2000young. (d) Top 6 most common pairwise social EOD motifs (Left) without collective sensing in non-competitive environments, vs. (Right) with collective sensing enabled. Higher occurrence of "silent" periods in one agent reveals emergent "freeloading" behavior during collective sensing. Social EOD motifs defined here as interaction between two agents within 15 cm for at least 4 timesteps. (e) SPI distributions shift upwards (equivalently, EOD rates lower) during collective sensing (CS) further supporting the emergence of "freeloading" strategies.
  • Figure 3: (a) Minimal social foraging assay with two agents, A and B. A is initialized within a fully-replenishing food patch, while B is randomly initialized within communication radius to A. (b) Example trajectories in different A/B relative dominance scenarios. (c) We vary the relative dominance levels of A/B, then compare the percentage of trials where B reaches the patch (100 runs). B performs better when it is more dominant. However, B's performance drops dramatically when agent A is removed, suggesting that there is a social component to foraging success. (d) Amount of food eaten by B per episode follows similar trends w.r.t. dominance. (e) B’s success is modulated by both dominance and starting location, indicating a social component to spatial foraging strategy pedrajaCollectiveSensingElectric2024.