DARPA Mathematical Challenge 13

BTUT

Planetary-scale multi-agent coordination through mean-field game theory. O(N) complexity. Provable convergence.

∂μ/∂t = -∇·(μ v[μ]) + σ²Δμ
Fokker-Planck Evolution
Live Convergence
Iteration: 0/30
Nash Gap: 1.0000
Agents: 100,000
O(N)
Complexity
20-30
Iterations to converge
10⁶+
Agents scalable
6
Application domains
Methodology

Mean-Field Game Theory

01

Population Distribution

Replace N² pairwise interactions with a continuous density function μ(x,t)

μ(x) = (1/N) Σᵢ δ(x - xᵢ)
02

Kernel-Weighted Sensing

Agents sense local environment through Gaussian kernel weighting

K(x,y) = exp(-||x-y||² / 2σ²)
03

Nash Equilibrium

System converges to equilibrium where no agent can unilaterally improve

ε = E[||uᵢ - u*||²] → 0
Theory

Understanding the Equation

The Fokker-Planck equation describes how a population of agents evolves over time. Each term has an intuitive physical meaning.

∂μ/∂t = −∇·(μ v[μ]) + σ²Δμ
Fokker-Planck Equation for Mean-Field Game Dynamics
Variables
μ
Population Density

Represents the distribution of all agents across the state space. Instead of tracking each individual agent, we track the "crowd" as a continuous density.

μ(x,t) = probability of finding an agent at position x at time t
t
Time

The simulation progresses in discrete iterations. Each step, agents update their strategies based on what they observe from the population.

∂μ/∂t = how the population changes over time
v[μ]
Velocity Field

The "direction" each agent wants to move based on what they observe. Agents move toward better strategies by sensing the local population density.

v[μ] = optimal direction given current population μ
σ
Exploration Noise

Controls how much agents "explore" vs "exploit". Higher σ means more random exploration; lower σ means agents stick closer to their current strategy.

σ² = variance of random perturbations (diffusion strength)
Operators
∇·
Divergence

Measures how much the population "flows out" of a region. Positive divergence means agents are leaving; negative means they're concentrating.

∇·F = ∂Fₓ/∂x + ∂Fᵧ/∂y (sum of partial derivatives)
Δ
Laplacian

Measures how "curved" the density is. Where density is peaked, the Laplacian is negative, causing spreading. This creates the diffusion effect.

Δμ = ∂²μ/∂x² + ∂²μ/∂y² (second derivatives)
The Two Forces
1
−∇·(μ v[μ]) — Drift

Agents move toward better strategies. This is the "exploitation" term — following the gradient toward Nash equilibrium.

2
+σ²Δμ — Diffusion

Random exploration spreads agents out. This is the "exploration" term — preventing premature convergence and ensuring global optimum discovery.

Intuition

Think of it like a crowd at a concert. The drift term is everyone trying to get closer to the stage (optimization). The diffusion term is people randomly moving around (exploration). Together, they find the best viewing spots where everyone is satisfied — the Nash equilibrium.

Convergence Metric

Nash Gap (ε)

ε = E[||uᵢ − u*||²]

The Nash gap measures how far agents are from the equilibrium strategy u*. When ε → 0, no agent can improve by changing their strategy unilaterally.

ε > 0.1 — Far from equilibrium
ε ~ 0.01 — Approaching equilibrium
ε < 0.001 — Converged (Nash equilibrium)
ε → 0
System converges in 20-30 iterations
Get Started

Build planetary-scale coordination

Connect real APIs, run live simulations, validate convergence.