Active Electrosensing and Communication in MARL-trained Weakly Electric Fish Collectives

AI summary

Overview

Research area: Computational neuroscience / neuroethology, multi-agent reinforcement learning (MARL), and active electrosensing (cs.NE; arXiv:2511.08436v2).

Technical level: Advanced. The paper combines biophysical electrosensory modeling, multi-agent PPO training, and recurrent-network analyses (PCA effective rank, linear decoding, partial least-squares correlation), so familiarity with reinforcement learning and neural population analysis helps.

Scope: The paper builds a simulated collective of weakly electric fish-like agents with biomimetic electroreceptors and trains them with MARL to forage socially, then uses in silico interventions and recurrent-dynamics analyses to probe how sensing and signaling shape group behavior.

What This Paper Is About

Studying how collective behavior arises from individual interactions is hard in real animals because simultaneous multi-brain recordings from freely interacting individuals are technically difficult and large behavioral assays are expensive and hard to control. Weakly electric fish are an appealing model because their electric organ discharges (EODs) simultaneously serve active sensing and communication, but no prior framework combined all three major mormyrid electroreceptor classes, active EOD generation, collective sensing, and closed-loop multi-agent decision-making. The authors build such a framework in simulation and test whether biophysically grounded sensing plus simple individual foraging incentives are sufficient to reproduce real fish behavior.

Key Contributions

1. A biophysically inspired MARL environment for fish-like agents. A custom 2D physics simulator handles mechanical and electrical interactions, including electric field generation, propagation, distortion by objects and walls, and sensing through three biomimetic electroreceptor classes (Mormyromasts, Ampullary receptors, Knollenorgans). Each agent has continuous forward translation and angular rotation actions plus discrete EOD emission and bite commands.

2. A training setup with no explicit rewards for social behavior. Agents use a GRU-based actor-critic policy trained with a multi-agent variant of Proximal Policy Optimization (PPO). Rewards are based on individual fitness: foraging success and asymmetric penalties during aggressive encounters, where the larger (dominant) agent incurs a smaller penalty than the smaller (subordinate) agent. No positive reward is given for communication, coordination, chasing, or aggression.

3. Validation against real fish behavior plus in silico causal interventions. The authors validate curvilinear homing and heavy-tailed inter-discharge interval (IDI) statistics against data from real fish, then perform sensor ablations, collective-sensing manipulations, EOD muting, and food-distribution changes.

4. Recurrent-dynamics analyses of social context. They show task-relevant variables and social context are encoded in RNN activity, and that pairs of interacting agents show proximity-dependent correlated latent dynamics.

Main Findings

• Curvilinear homing. In a 70 cm × 70 cm arena with one fixed always-emitting target agent, a homing agent initialized at random positions across 100 trials followed curved paths, whereas a non-electric ray-casting baseline that received direct line-of-sight distance and object-type cues followed straight lines. Error angles (difference between the agent's heading and the local electric field direction of the target) showed a bimodal distribution near parallel and anti-parallel alignment with field lines and progressive alignment over successful episodes, qualitatively resembling statistics reported for real fish.

• Heavy-tailed EOD statistics. In a two-agent wide assay (160 cm × 40 cm arena, 100 evaluation episodes), inter-discharge intervals during foraging were heavy-tailed, as in real fish, and both admitted power-law-like fits over the same range, though the exponents did not quantitatively match.

• Context-dependent EOD modulation. EOD rate varied systematically with distance to the nearest conspecific, distance to the closest food item, and distance to the nearest arena wall. Peri-event analyses showed a sharp transient increase around eating, a build-up before and decline after biting another fish, and elevation around times the agent was bitten.

• Emergent dominance-like asymmetries. With size and speed heterogeneity, larger agents directed more biting toward smaller conspecifics, win ratio increased with body size, and larger agents consumed more food per episode. Food-consumption inequality (Theil index) scaled with body-size inequality across groups. Body size tended to correlate positively with EOD emission rate, and EOD rate advantage weakly tracked size advantage in close-range pairs (≤ 10 cm).

• Short-range sensors support foraging. Ablating either Mormyromast or Ampullary receptors reduced food consumption in both a solo assay and four-agent groups, indicating effects on individual foraging that persist without conspecifics.

• Long-range Knollenorgans shape social organization. Knollenorgan ablation left total food consumption largely intact but increased EOD emission probability and biting and reduced inter-agent spacing.

• Collective (conspecific-EOD) sensing matters. Within the Mormyromast channel, removing the conspecific-EOD image (self-image-only) drastically reduced food consumed, while removing the self-EOD image (cons-image-only) did not show a significant foraging effect. EOD emission probability was lower in self-image-only agents and higher in cons-image-only agents.

• EOD muting. Silencing EOD emission in one of two agents reduced total food consumption and increased mean nearest-neighbor distance.

• Environmental structure. Patchy food distributions yielded substantially more food per episode than uniform layouts, with no significant difference in inequality. In a patch-number sweep from 1 to 5 patches, per-fish intake fell progressively in the iso-food series (total food constant) relative to the unconstrained series, and inequality fell with increasing patch count in both series, somewhat more steeply in the unconstrained condition.

• Dyadic competition depends on size and signaling. In 150 cm × 150 cm arenas with a resident on a replenishing patch and an intruder initialized away from it (100 trials per size condition), dominant intruders approached directly and displaced residents, same-size pairs often shared access, and subordinate intruders approached more indirectly. Intruder B consumed significantly more food when A was present than when foraging alone, while resident A's intake was not significantly affected by B's presence. With a rule-based patch-confined random-walker "bot" in place of the resident, B reached the patch more often when the bot was frozen than when moving, and B's patch-reach rate was generally higher when the bot emitted EODs at any nonzero rate than when silent, peaking at intermediate emission rates. In separate 70 cm × 70 cm arenas with uniform food, larger agents were more likely to be chasers than chased, biters tended to be larger than victims, biting peaked at approximately 40% of normalized interaction duration, and confronting interactions had the highest biting rate (≈ 50%), followed by chasing, fleeing, and unaligned interactions.

• Low-dimensional recurrent dynamics. PCA of four-agent patchy foraging episodes gave an effective rank of D_eff = 81.3 (304 PCs for 90% variance), well below the nominal D_H = 512 hidden-state size. Effective dimensionality increased with the number of interacting agents when Knollenorgans were intact, but stayed flat near the solo baseline when Knollenorgans were ablated.

• Field-line geometry is represented. A linear decoder recovered the Knollenorgan-defined error angle with substantially higher performance than the straight-line angle to the target. In the two-agent wide assay, nearest-agent distance and Knollenorgan features (presence and count) were the top-decoded variables, followed by log Mormyromast field magnitude and wall distance; food distance and food bearing tended to decode better when food was in sensing range.

• Correlated latent structure between agents. Partial least-squares correlation first-component correlation decreased across conspecific-distance bins (≤ 10, 10–100, > 100 cm). The number of statistically significant shared dimensions was similar for pairs within 100 cm (not significant between the ≤ 10 and 10–100 cm bins) but collapsed to near zero beyond 100 cm.

• RNN power effects. Mean wideband RNN power was significantly lower at timesteps when a conspecific was nearby, and subordinate (smaller) agents carried significantly higher mean wideband power than dominant (larger) agents.

Methodology in Plain English

The authors built a custom 2D simulator in which each agent is a circular particle that can move forward, rotate, emit EODs, and bite. Electric fields in the arena come from actively controlled EOD sources on agents and from intrinsic electric sources associated with agents and prey, and these fields can be distorted by nearby objects, so agents sense how their own and others' discharges are transformed by the shared space. Each agent carries three receptor types modeled on mormyrid fish: Mormyromasts (short-range active sensing of prey within ≤ 5 cm and agents within ≤ 10 cm, plus a conspecific-EOD "cons-image"), Ampullary receptors (passive, short-range sensing of prey within ≤ 4 cm and conspecifics within ≤ 8 cm), and Knollenorgans (long-range, ≤ 100 cm, detection of conspecific EOD pulses that identify the direction and size of emitting agents but do not directly encode distance).

Each agent's policy is a GRU-based actor-critic network with feedforward action and value heads, trained with a multi-agent variant of PPO. A scalar "size" parameter scales maximum speed and is given as fixed context. Multiple independent seeds were trained, a controlled behavioral assay was run on each, and one representative converged policy was selected using an ethologically motivated criterion balancing group foraging success, size-dependent resource asymmetry, and broad agent participation. Post-training analyses used that policy across several group sizes, arena and food configurations, and sensory perturbations. Statistical comparisons included one-vs-control Dunnett tests for receptor ablations and exploratory Mann-Whitney or Wilcoxon tests where appropriate, with PLSC significance assessed via a circular-shift null.

Why This Matters

Impact on research. The paper instantiates a "virtual neuroscience" approach for electrosensory collective behavior: it provides qualitative validation against established biological data while enabling interventions that are impractical or impossible in vivo. Because every agent's sensory inputs, neural activity, and actions are fully observable and the experimenter controls sensing, signaling, and reward structure, it enables mechanistic analysis not yet accessible in real animals. It also claims to be the first framework combining all three major mormyrid electroreceptor classes, active EOD generation, collective sensing, and closed-loop multi-agent decision-making, producing electrocommunication patterns consistent with real fish from individual fitness incentives alone. The authors state their goal is not a complete biophysical replica and that they do not claim the discovered mechanisms are uniquely required in biological animals.

Real-world applications (framed by the authors as implications, not demonstrated deployments):

• Generating testable predictions for ablation and competition experiments in real weakly electric fish, such as whether blocking Knollenorgan-mediated signaling increases agonistic interactions and reduces social spacing without impairing individual foraging.
• Guiding future data collection in systems where simultaneous behavioral tracking, EOD assignment, and multi-animal neural recordings remain technically challenging.
• Providing an engineering benchmark for multi-agent systems that must sense through a physically grounded, actively generated modality rather than simplified sensory representations.
• Informing broader questions about embodied social intelligence and the consequences of disrupted sensing or communication for collective behavior in other multi-agent sensory systems.

Industry relevance. The framework's ingredients — biophysically grounded sensing, decentralized policies trained with individual incentives, emergent role asymmetries, and correlated latent dynamics between interacting agents — map onto problems in multi-robot coordination, distributed sensing networks, and multi-agent systems where agents must communicate while simultaneously sensing through a shared physical medium. The paper does not report commercial deployments or benchmarks; relevance here is conceptual rather than demonstrated.

Future Directions

• Experimental tests of the Knollenorgan prediction. The authors propose that selectively blocking Knollenorgan-mediated signaling in real fish groups should increase agonistic interactions and reduce social spacing without substantially impairing individual foraging efficiency.
• Dyadic displacement predictions. They predict that larger intruders should systematically displace smaller residents, with displacement probability scaling with the size difference, and that intruders should reach occupied patches more readily when the resident is stationary or actively emitting.
• Subordinate EOD suppression. The agents did not show a full cessation of EOD emission by subordinate individuals as observed in real fish, though a size-dependent reduction in EOD emission rate during interactions appeared. Whether learned suppression emerges under stronger social pressure or richer communication structure is posed as a direct test for future work.
• Data collection and generalization. The model is intended to precede and guide future data collection rather than replace it, given that simultaneous behavioral tracking, EOD assignment, and multi-animal neural recordings remain technically challenging; extending the framework to other social animals and multi-agent sensory systems is also raised.

Target Audience

Neuroethologists and experimentalists who study weakly electric fish and social behavior; computational neuroscientists working on recurrent population dynamics, latent-variable comparisons across individuals, and "virtual neuroscience" models; and multi-agent reinforcement learning researchers interested in biophysically grounded sensing, emergent communication, and decentralized coordination under individual incentives. Readers without a background in reinforcement learning or neural population analysis will find the modeling and decoding sections demanding.

Authors’ abstract

How complex collective behavior emerges from individual interactions is a fundamental scientific question, but experimental cost and difficulty of simultaneous multi-brain recordings limit direct study in animals. Here we introduce a novel computational framework modeling weakly electric fish-like agents with biophysically inspired electrosensing and actuation, trained to forage collectively via multi-agent reinforcement learning (MARL). Trained agents reproduce hallmarks of real fish, including curvilinear homing trajectories and heavy-tailed electric organ discharge (EOD) interval statistics, while exhibiting emergent active sensing, social foraging, dominance-like asymmetries, and aggression. We perform in silico interventions including sensor ablations, EOD silencing, and food distribution changes to identify causal drivers of social foraging. Analyses of recurrent neural dynamics further show robust encoding of task-relevant variables and social context. Our work has broad implications for the neuroethology of weakly electric fish and other social animals where extensive multi-individual neural recordings, and thus traditional data-driven modeling, remain challenging.