Summary

Three distinct deep learning approaches for MFG/MFC, each with different trade-offs for implementation:

Method 1: Direct SGD for MKV Control (Algorithm 1). Parameterize the control alpha as a neural network alpha_theta(t, x). Simulate N interacting particles with empirical mean-field distribution. Discretize time. Minimize the cost functional J^{N,Delta t}(theta) directly via SGD — no PDE, no FBSDE, no optimality conditions needed. This is the simplest to implement. Applied to a price impact model with transaction costs (Section 2.2): traders with quadratic transaction costs c_alpha/2 * alpha^2 and holding costs c_X/2 * X^2 interact through aggregate trading rate. Parameters: N=2000 particles, N_T=50 time steps, sigma=0.5, T=1.

Method 2: Deep BSDE Shooting for MKV FBSDEs (Algorithm 2). For MFG equilibria characterized by coupled FBSDEs, parameterize the initial value Y_0 = y_{0,theta} and the volatility of Z = z_omega as neural networks. Simulate forward, compute terminal mismatch loss J_{FBSDE} = E|Y_T - G(X_T, L(X_T))|^2. SGD on (theta, omega). Applied to systemic risk MFG (Section 3.2) — the same inter-bank model used in the DFP papers. Handles common noise by conditioning on the filtration of W^0.

Method 3: Deep Galerkin Method for Mean Field PDEs (Algorithm 3). For MFG PDE systems (coupled KFP + HJB), parameterize density m = m_{theta_1} and value u = u_{theta_2} as two neural networks. Loss = L^{KFP}(m,u) + L^{HJB}(m,u) where each term is the L^2 PDE residual plus boundary condition penalties (eqs. 17-18). Collocation points sampled uniformly. Applied to crowded trade model with transaction costs (Section 4.2) where kappa > 0 is the quadratic transaction cost parameter.

Key Contributions

  • Unified survey of three complementary deep learning approaches for MFG/MFC
  • All three illustrated on financial applications with transaction costs or price impact
  • Method 1 requires no knowledge of optimality conditions — just simulate and optimize
  • Method 2 extends the deep BSDE method to McKean-Vlasov FBSDEs
  • Method 3 (deep Galerkin method) is the most general but hardest to tune

Methodology

The methods target different formulations of the same problem: Method 1 works at the control level (no PDE), Method 2 at the FBSDE level (probabilistic), Method 3 at the PDE level (analytical). For implementation, Method 1 is simplest (just differentiate the cost), Method 2 leverages the BSDE structure for efficiency, Method 3 is most flexible but requires careful tuning of loss weights C^{KFP}, C_0^{KFP}, C^{HJB}, C_T^{HJB}.

Key Findings

  • Method 1 (direct SGD) matches semi-explicit ODE solutions for price impact model
  • Method 2 (BSDE shooting) approximation is better for X^i than Y^i (errors accumulate in forward Y equation)
  • Method 3 (DGM) loss weight selection is critical — poor weights lead to trivial solutions
  • Common noise makes all methods harder; Method 2 handles it by conditioning
  • Transaction cost examples appear in both Methods 1 and 3

Important References

  1. Probabilistic Theory of Mean Field Games with Applications I — Carmona, Delarue (2018), theoretical foundation
  2. Mean Field Games and Systemic Risk — Carmona, Fouque, Sun (2015), the systemic risk model used as benchmark
  3. Convergence Analysis of Machine Learning Algorithms for MFG and MFC — Carmona, Lauriere (2019/2021), convergence proofs

Atomic Notes


paper