Reproduce “Emergent behavior and neural dynamics in artificial agents tracking turbulent plumes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Satpreet Harcharan Singh, Floris van Breugel, Rajesh P. N. Rao, Bingni Wen Brunton
The authors train recurrent neural network agents with reinforcement learning to locate odor sources in two-dimensional simulations of intermittent plumes and changing wind. Agents use recent odor and wind signals to choose movement. They generally outperform fixed-history feedforward networks, especially when wind switches. Their movement aligns more with plume centerlines than current wind, suggesting a hypothesis for insect experiments.
Key insights
- DRL-trained RNN agents autonomously develop complex, stereotyped behavioral modules (tracking, recovering, lost) for plume tracking that qualitatively resemble those observed in flying insects.
- Agents prioritize tracking the plume's historical centerline rather than the instantaneous wind direction, especially in dynamic wind conditions, suggesting a testable hypothesis for biological systems.
- Recurrence provides a critical advantage for plume tracking, enabling the network to maintain dynamic, context-dependent memory that outperforms fixed-length sensory histories in complex environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.