Digital Lab
The simulation arena that tests every rule under identical friction.
The Digital Lab models a bounded 2D world with noise, turbulence, and energy costs. Each agent is intentionally "dumb" and must decide using only local sensing and limited state.
Environment
- Static inverse-square field or intermittent plume puffs.
- Wind advection + diffusion for turbulent regimes.
- Sensor noise and optional boundary conditions.
Agents
- Energy-limited motion with sensing costs.
- Local sensing radius and neighbor awareness.
- Finite-state controllers for intermittent search.
Metrics
- Time to first find, time to all targets.
- Energy per target, success rate.
- Variance across Monte Carlo trials.
Rule Families
Four behavioral regimes, one switching logic.
Random walk vs. Levy flight
Brownian motion oversamples local space in sparse domains. Levy steps inject rare, long jumps that break recurrence, but only help in specific density regimes.
Relocation + detection phases
Fast, blind relocation alternates with slow, high-sensitivity detection. The switching ratio is the main control variable.
Infotaxis vs. surge-cast
Infotaxis moves by information gain; surge-cast approximates it with a two-state controller driven by odor hits and wind direction.
Stigmergy + quorum
Indirect pheromone memory expands coverage, while quorum thresholds stabilize lock-in when a target is confirmed.
Law of Search
The switching boundary between exploration and exploitation.
When information gain per unit energy crosses a threshold, behavior flips from coverage to capture.
Confidence Ratio
chi = (signal / noise) * (1 / rho_est) * energy
Low chi means low information density. High chi means strong evidence and nearby targets.
Simulator
Run the leaderboard in your browser.
The in-browser Digital Lab runs a fast, simplified version of the simulator so you can compare strategies directly on GitHub Pages. For full fidelity and plots, use the Python version in sim/.
| Strategy | Steps to First | Steps to All | Success Rate | Energy |
|---|
Avg Steps to First
Success Rate
Live agent animation
Preview a single strategy in motion. This is a lightweight visual, not the full leaderboard.
cd sim
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python run_leaderboard.py --trials 20 --plume turbulent --save-plotsFull plots and higher-fidelity runs in sim/figures/.
References
Key sources behind the unified theory.
- Vergassola, Villermaux, Shraiman. Infotaxis as a strategy for searching without gradients. Nature (2007). DOI: 10.1038/nature05464.
- Benichou et al. Intermittent search strategies. Rev Mod Phys (2011). DOI: 10.1103/RevModPhys.83.81.
- Lomholt et al. Levy strategies in intermittent search processes are advantageous. PNAS (2008). DOI: 10.1073/pnas.0803117105.
- Palyulin, Chechkin, Metzler. Levy flights do not always optimize random blind search for sparse targets. PNAS (2014). DOI: 10.1073/pnas.1320424111.
- Weisstein. Polya's Random Walk Constants. MathWorld.
- Wechsler, Bhandawat. Odor-modulated locomotion in insects. J Exp Biol (2023). DOI: 10.1242/jeb.200261.
- Verano, Panizon, Celani. Olfactory search with finite-state controllers. PNAS (2023). DOI: 10.1073/pnas.2304230120.
- Heylighen. Stigmergy as a universal coordination mechanism I. Cognitive Systems Research (2016). DOI: 10.1016/j.cogsys.2015.12.002.
- Kennedy, Eberhart. Particle swarm optimization. Proc. IEEE ICNN (1995). DOI: 10.1109/ICNN.1995.488968.
- Scheidler et al. The k-unanimity rule for self-organized decision-making in swarms of robots. IEEE Trans Cybernetics (2016). DOI: 10.1109/TCYB.2015.2429118.
Full citations and context are in the whitepaper.