MARL-Trained Virtual Weakly Electric Fish Emerge Realistic Social Behaviors
The Kempner Institute released a study utilizing multi-agent reinforcement learning (MARL) to train virtual weakly electric fish, detailed in the paper "Active Electrosensing and Communication in MARL-trained Weakly Electric Fish Collectives." This system provides a fully observable and controllable computational testbed for studying social behavior, with the model and supplementary materials open-sourced.
已确认
- 研究搭建了受弱电鱼启发的多智能体强化学习框架，每个RNN代理可移动、发出电脉冲和撕咬，以觅食奖励为主，对被咬施加惩罚。
- 物理引擎模拟了鱼的运动与动态电场环境，每条鱼通过EOD脉冲产生电场，周围个体、食物和边界墙均会造成扰动。
- 模型借鉴真实弱电鱼设置M、A、K三种受体，M和A负责短距离感知，K负责长距离社交信号感知。
- 在没有显式社交规则的情况下，代理自发涌现竞争行为，包括追逐和撕咬，并形成明确的体型等级秩序：更大的鱼稳定获胜，体型差异越大食物消费不平等越严重。
- 仿真鱼的EOD统计特征与生物数据相似，包括放电间隔呈长尾分布、觅食时放电率上升、靠近更大鱼时放电受抑制。
- 电场感知使回家任务产生更平滑、弯曲且沿场几何结构前进的路径，优于ray-casting对照。
- RNN代理内部表征高度压缩（有效秩远低于网络宽度），能稳定编码任务与社交变量，如沿电场线回家角度和社交距离，近距离时神经活动对齐更强。
为什么重要
- 该框架证明复杂社交行为可从简单的觅食奖励与物理交互中涌现，无需预设社会规则，为理解真实鱼群提供了可证伪的假设生成工具。
- 虚拟系统支持in vivo难以实现的干预实验，有助于优先排序关于真实鱼群的科学假设。
2026-07-29 ~ 2026-07-29 · 11 related posts
Primary sources
- Kempner Institute open-sources the weakly electric fish collective model and paper materials — tweetsatpreet ·
- MARL-trained Virtual Fish Exhibit Emergent Social Behaviors Like Real Animals — tweetsatpreet ·
- Multi-agent fish simulation uses RNN agents with electric sensing and biting — tweetsatpreet ·
- [source] MARL-trained Virtual Fish Exhibit Emergent Social Behaviors Like Real Animals — tweetsatpreet · 2026-07-29
- A physics engine simulates fish movement and a dynamic electric landscape — tweetsatpreet · 2026-07-29
- Simulated electric fish use M, A and K receptors for short- and long-range sensing — tweetsatpreet · 2026-07-29
- [source] Multi-agent fish simulation uses RNN agents with electric sensing and biting — tweetsatpreet · 2026-07-29
- Electric-field sensing makes simulated fish match real discharge and homing patterns — tweetsatpreet · 2026-07-29
- Aggression emerges in fish simulation without explicit social rules — tweetsatpreet · 2026-07-29
- Bigger fish win more often and eat more in a size-ordered simulation — tweetsatpreet · 2026-07-29
- Two-agent fish simulation shows size-based pecking order and unequal food sharing — tweetsatpreet · 2026-07-29
- Fish RNNs learn compact codes that better decode homing angles and social distance — tweetsatpreet · 2026-07-29
- Virtual fish give researchers a fully observable testbed for social behavior hypotheses — tweetsatpreet · 2026-07-29
- [source] Kempner Institute open-sources the weakly electric fish collective model and paper materials — tweetsatpreet · 2026-07-29