(5/5)
Even better: these serial steps can be vectorized across many models. Used already for traffic, markets, predator–prey & foraging. More domains coming soon 🙌
Thanks to collaborators & @dbi2program.bsky.social .
(4/5)
Scaling ABMs hits a wall at conflict management (inherently serial). ABMax solves this with two JIT callbacks: Rank-Match (fast) and Sort-Count-Iterate (modeler-friendly).
(3/5)
The API blends OOP with JAX-style functional programming: define agents, policies, interactions—and operate on agent sets (sort, select, update) cleanly.
(2/5)
JAX already makes simulations scale effortlessly with JIT + vectorization. ABMax brings the same ease to agent-based models, without giving up Pythonic design.
(1/5)
Another preprint 📄: ABMax — a JAX-based agent-based modeling framework.
arxiv.org/abs/2508.16508
github.com/i-m-iron-man...
(5/5)
We also use a single-brain/multi-body evolution setup, letting policies learn interactions, not just behaviors. Built with ABMax (JAX) (github.com/i-m-iron-man...).
Thanks to collaborators at @dbi2program.bsky.social.
(4/5)
Inside the agents’ brains, some units track internal resource levels. Forcing these units to “think” they’re hungry speeds up swarming—echoing urgency-gating mechanisms in neuroscience.
(3/5)
Our model shows an emergent rule: hungrier agents swarm more. Well-fed agents disperse. This mirrors the asset-protection principle from social foraging theory.
(2/5)
Seeing another forager is ambiguous: it might signal nearby food—or a depleted patch. So agents must decide: follow others or avoid them?
(1/5)
New preprint out 🎉
We study how swarming emerges in multi-agent patch foraging when agents only have partial, first-person sensing.
arxiv.org/pdf/2510.18886
Had an amazing experience at the Mechanistic basis of Foraging 2025 conference.
