Chiyoda: Entropy-Guided Information Control for Hazard-Coupled Pedestrian Evacuation
- Date
- 30 Apr 2026
- Authors
- Gabriel Ong Zhe Mian
- Zenodo
- 10.5281/zenodo.19905069
https://zenodo.org/records/19905070 - Resource type
- Publication
- License
- CC-BY-4.0
- Version
- v1
- GitHub Repository
- https://github.com/gongahkia/chiyoda
I recently published the first version of Chiyoda on Zenodo.
Lore
As my first self-directed "proper" research paper, getting it into a form that could be read, cited (and possibly criticised) felt very different from shipping a normal
software project.
While I originally began Chiyoda as a side-project in 2024, this process of drafting a paper about it really forced me to scope down its claims to those sustantiated by clear evidence. Presently (in April 2026) at least, Chiyoda is a research toolkit for studying emergency evacuation as an information problem.
Problem
Most evacuation systems implicitly make the assumption that communicating the hazard will always be helpful, predicated on the notion that better warnings informs the crowd to effect better decisions.
What Chiyoda does different
In Chiyoda, agents do not move through a perfect map of the world. Instead, they carry pre-existing beliefs (about exits, hazards, congestion, and danger etc.) and can actively observe their surroundings, receive messages (via beacons), hear from emergency responders and pass distorted local gossip to each other.
These imperfect notions then direct how the agents route through the world.
That distinction - that a message can both be correct and nonetheless remain "risky" information was what interested me in 2024 and what continues to drive the writing of this paper in 2026.
TlDR: When does communication of a risk improve belief and safety holistically, and when does it creates harmful convergence (in a physical medium)?
Thoughts
Result reproducibility was a way larger part of drafting the research paper than I initially expected.
In terms of possible angles for further unexplored scope, LLM-bounded and generated evacuation messages remains an area the paper touches on only briefly.
Closing
Either way, v1 of Chiyoda is the iteration I'm happiest with right now. Really grateful for the experience, and I also extend my thanks to anyone who does take the time to read it and give me their feedback.
If this sounds interesting to you, feel free to check it out on Zenodo at 10.5281/zenodo.19905070