Recurrent Neural Cellular Automata with Self-Attention for Multi-agent System
Keywords: neural cellular automata, complex system, self-organization, multi-agent system
TL;DR: A neural cellular automata model that effectively discovers the dynamic and stochastic local interaction with limited data
Abstract: Many-agent systems, such as epidemic spread, rumor propagation through crowd, prey-predator model, and forest fire, exhibit complex global dynamics originated from local, relatively simple, and often stochastic interactions between agents. Despite significant advancements in predictive modeling through deep learning, such interactions among many agents have rarely explored as a specific domain for predictive modeling. We present Recurrent Attention-based Neural Cellular Automata (RA-NCA), to effectively discover the local stochastic interaction by associating the temporal information between neighboring agents in a permutation-invariant manner. RA-NCA exhibits the superior generalizability across various agent configurations (i.e., spatial distribution of agents), data efficiency and robustness in extremely data-limited scenarios even with the presence of stochastic interactions, and scalability through spatial dimension-independent prediction. We compare and evaluate RA-NCA with other NCA networks and scene prediction networks in the three synthetic multi-agent systems with thousands of agents, such as forest fire, host-pathogen, and stock market models.
Supplementary Material: zip
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8157
Loading