Slide 1

Slide 1 text

Mean Field Games with Applications 1

Slide 2

Slide 2 text

1 !2

Slide 3

Slide 3 text

Work of 2018-2019 Social Network Engineering Gradient flow !3

Slide 4

Slide 4 text

Mean-field games (MFG) framework Introduced by P.-L. Lions, J.-M. Lasry in the mathematics community and M. Huang, P. Caines, R. Malham´ e in the engineering one to model huge populations of identical, rational agents that play non-cooperative di↵erential games. In a PDE form, MFG resemble macroscopic equations from statistical physics in game-theoretic/optimal control framework. Yield near optimal solutions for huge populations of optimizing agents. Thus, tremendous simplification in practical problems. engineer field !4

Slide 5

Slide 5 text

Nash equilibrium in a differential game A continuum of players with a distribution density ⇢(x, t) play a non-cooperative di↵erential game. An individual agent faces an optimization problem (x, t) = inf E t Z 0 L(Xs, vs, ⇢(·, s), s)ds + 0(X0, ⇢(·, 0)), (9) where dXs = vsds + p 2 dWs, XT = x, and L is the Lagrangian. In Nash equilibrium, each agent cannot do better unilaterally. Optimal strategies are given by the value function, , that solves the corresponding Hamilton-Jacobi equation. The density, ⇢(x, t), evolves according to optimal actions of agents. Thus, the Fokker-Planck equation. !5

Slide 6

Slide 6 text

The PDE formulation A typical time-dependent MFG system (derived from (9)) has the form 8 > > > > > < > > > > > : t 2 + H (x, Dx (x, t), ⇢(x, t), t) = 0 ⇢t 2 ⇢ divx (⇢DpH (x, Dx (x, t), ⇢(x, t), t)) = 0 ⇢ 0, R Td ⇢(x, t)dx = 1, t 2 [0, T] ⇢(x, T) = ⇢T (x), (x, 0) = 0(x, ⇢(x, 0)), x 2 Td . (10) Given: H : Td ⇥ Rd ⇥ X ⇥ R ! R is a periodic in space and convex in momentum Hamiltonian, where X = R+ or X = L 1(Td ; R+) or X = R+ ⇥ L 1(Td ; R+). H = L ⇤ is a di↵usion parameter 0 : Td ⇥ X ! R, ⇢T : Td ! R given initial-terminal conditions Unknowns: , ⇢ : Td ⇥ [0, T] ! R above !6

Slide 7

Slide 7 text

Classification of MFG Mean-field interaction or non-linearity is the dependence of H, L on the density, ⇢. Local interaction if H, L depend on ⇢ locally. For instance, H(x, p, ⇢, t) = ˜ H(x, p, t) ln ⇢ H(x, p, ⇢, t) = ˜ H(x, p, t) ⇢ Non-local interaction if H, L depend on ⇢ non-locally. For instance, H(x, p, ⇢, t) = ˜ H(x, p, t) F 0 @ x, Z Td K(x, y)⇢(y)dy 1 A Congestion interaction if we penalize moving in dense area. For instance (soft congestion) H(x, p, ⇢, t) = ⇢↵ ˜ H ✓ x, p ⇢↵ , ⇢, t ◆ , ↵ > 0 !7

Slide 8

Slide 8 text

U(t, µ) = inf ⇢2P(Td) F(⇢) + dW (⇢, µ)2 2t Hopf-Lax on density space E.g. (Burgers’) Hamilton-Jacobi on density space Characteristics on density space (Nash equilibrium in mean field games) @tµs + r · (µs r s) = 0 @t s + 1 2 (r s)2 = 0 Math Review: mean field games @ @t U(t, µ) + 1 2 Z Td (r µ U(t, µ))2µ(x)dx = 0, U(0, µ) = F(µ) !8

Slide 9

Slide 9 text

Density Control: Optimal Transport dW (⇢, µ)2 = inf ⇢s,us n Z 1 0 Z R u2 s ⇢sdxdt: @s⇢s + r · (⇢sus) = 0, ⇢0 = ⇢, ⇢1 = µ o 8 What happens if the total population density is not normalized/balanced? 9

Slide 10

Slide 10 text

works !10

Slide 11

Slide 11 text

Change delta F to negative delta F mean field game !11

Slide 12

Slide 12 text

| Z µ Udx|2 !12

Slide 13

Slide 13 text

!13

Slide 14

Slide 14 text

0 1 0 Density (7(x; t)) Time (t) 0.5 0.5 0.5 Space (x) 1 0 1 14

Slide 15

Slide 15 text

Unnormalized L1 Wasserstein metric Let p = 1: UW1(µ0, µ1) = inf v,f(t) n Z 1 0 Z ⌦ kvkµdxdt + 1 ↵ Z 1 0 |f(t)|dt · |⌦|: @tµ + r · (µv) = f(t) o . 15 15

Slide 16

Slide 16 text

Time independent solution Denote m(x) = Z 1 0 v(t, x)µ(t, x)dt, with the fact Z 1 0 f(t)dt = c = 1 |⌦| ⇣ Z ⌦ µ1(x)dx Z ⌦ µ0(x)dx ⌘ . then by Jensen’s inequality and integrating the time variable t, we obtain UW1(µ0, µ1) = inf m n Z ⌦ km(x)kdx + 1 ↵ Z ⌦ µ0(x)dx Z ⌦ µ1(x)dx : µ1(x) µ0(x) + r · m(x) = 1 |⌦| ⇣Z ⌦ µ1(x)dx Z ⌦ µ0(x)dx ⌘o . 16 From this equation, we see the momentum is independent of alpha. 16

Slide 17

Slide 17 text

Closed form solution In one space dimension on the interval ⌦ = [0, 1], the L1 unnormalized Wasserstein metric has the following explicit solution: UW1(µ0, µ1) = Z ⌦ Z x 0 µ1(y)dy Z x 0 µ0(y)dy x Z ⌦ (µ1(z) µ0(z))dz dx + 1 ↵ ⇣ Z ⌦ µ1(z)dz Z ⌦ µ0(z)dz ⌘ . 17 If the total masses are equal, the additional term disappears. 17

Slide 18

Slide 18 text

Unnormalized L2 Wasserstein metric Let p = 2: UW2(µ0, µ1) = inf v,f(t) n Z 1 0 Z ⌦ kvk2µdxdt + 1 ↵ Z 1 0 f(t) 2dt · |⌦|: @tµ + r · (µv) = f(t) o . 22 18

Slide 19

Slide 19 text

Minimizer system The minimizer (v(t, x), µ(t, x), f(t)) for UOT problem satisfies v(t, x) = r (t, x), f(t) = ↵ 1 |⌦| Z ⌦ (t, x)dx, and 8 > > > > < > > > > : @tµ(t, x) + r · (µ(t, x)r (t, x)) = ↵ 1 |⌦| Z ⌦ (t, x)dx @t (t, x) + 1 2 kr (t, x)k2 = 0 µ(0, x) = µ0(x), µ(1, x) = µ1(x). 23 19

Slide 20

Slide 20 text

Unnormalized Monge-Ampere equation Denote (x) = 1 2 kxk2 + (0, x), Following the Hopf-Lax formula, the minimizer of unnormalized OT satisfies µ(1, r (x))Det(r2 (x)) µ(0, x) =↵ Z 1 0 Det ⇣ tr2 (x) + (1 t)I ⌘ · Z ⌦ ⇣ (y) kyk2 2 + tkr (y) yk2 2 ⌘ Det ⇣ tr2 (y) + (1 t)I ⌘ dydt. 24 20

Slide 21

Slide 21 text

Unnormalized Kantorovich problem 1 2 UW2(µ0, µ1) 2 = sup n Z ⌦ (1, x)µ(1, x)dx Z ⌦ (0, x)µ(0, x)dx ↵ 2 Z 1 0 ⇣ Z ⌦ (t, x)dx ⌘2 dt o where the supremum is taken among all : [0, 1] ! ⌦ satisfying @t (t, x) + 1 2 kr (t, x)k2  0. 25 This is connected to one of UCLA’s fast algorithms for OT by M. Jacobs and F. Leger, 2019. 21

Slide 22

Slide 22 text

Primal-dual method !22

Slide 23

Slide 23 text

Algorithm for MFG !23

Slide 24

Slide 24 text

Example I 29 24 OT

Slide 25

Slide 25 text

Example II 30 25

Slide 26

Slide 26 text

Example 31 26

Slide 27

Slide 27 text

Generalized unnormalized OT In fact, when p=2, the minimization has the unconstrained optimization formulation: Models suggested by Chen, Georgiou, Ling, Tannenbaum, 2017 for p=1,2. where Lµ = r · (µr) !27

Slide 28

Slide 28 text

Unconstrained formulation The corresponding gradient descent flow becomes !28

Slide 29

Slide 29 text

Gradient flow algorithm !29

Slide 30

Slide 30 text

Examples of unnormalized OT in 1D f(t,x) f(t) !30

Slide 31

Slide 31 text

Examples of unnormalized OT in 2D f(t,x) !31

Slide 32

Slide 32 text

Mean field game on manifolds Here we replace the domain of MFG from torus strategy set to a manifold M, point clouds, and graphs (discrete strategy set) U(t, µ) = inf v,µ,f Z 1 0 Z M kv(t, x)kpµ(t, x) F(µ(t, ·))dxdt + G(µ(0, ·)) such that the dynamical constraint is satisfied @tµ + r · (µv) = 0, µ(1, x) = µ1(x) Consider !32

Slide 33

Slide 33 text

Optimal transport on Manifold !33

Slide 34

Slide 34 text

MFG in 1D via unconstrained optimization Evolution of density with Phi(0,x) fixed !34

Slide 35

Slide 35 text

Classical MFG examples via unconstrained optimization The evolution of density Quadratic interaction energy with linear periodic potential energy We test our algorithms using examples in Achdou, Capuzzo-Dolcetta, Mean Field Games: Numerical Methods, 2009. !35

Slide 36

Slide 36 text

Classical MFG examples via unconstrained optimization The evolution of density: quadratic interaction energy with linear periodic potential energy !36

Slide 37

Slide 37 text

Spectral methods in nonlocal MFG Goal: Solve MFG with nonlocal interactions 8 > < > : t + D + H(x, r ) = R ⌦ K(x, y)⇢(y, t)dy, ⇢t + D⇤⇢ div(⇢rpH(x, r )) = 0, (x, 0) = 0(x), ⇢(x, T) = ⇢T (x). (13) Above, K is a monotone symmetric interaction kernel, D is a monotone linear operator (e.g. D = ). Strategy: Model the interaction in a Fourier space. Specifically, given a basis {x 7! fi (x)}i expand the kernel K(x, y) = X i,j kij fi (x)fj (y). !37

Slide 38

Slide 38 text

Furthermore, Z ⌦ K(x, y)⇢(y, t)dy = X i ai (t)fi (x), where ai (t) = X j kij Z ⌦ ⇢(y, t)fi (y)dy. Hence, (13) can be written as 8 > > > > < > > > > : t + D + H(x, r ) = P i ai (t)fi (x), ⇢t + D⇤⇢ div(⇢rpH(x, r )) = 0, ai (t) = P j kij R ⌦ ⇢(y, t)fi (y)dy, 8i, (x, 0) = 0(x), ⇢(x, T) = ⇢T (x). (14) The problem !38

Slide 39

Slide 39 text

New variational principle: System (14) is equivalent to inf a ⇢ a · K 1 a 2 Z ⌦ a(x, T)⇢T (x)dx s.t. (16) holds , (15) where K = (kij ), and a is the viscosity solution of ( t + D + H(x, r ) = P i ai (t)fi (x), (x, 0) = 0(x). (16) Advantages: (15) is a convex optimization problem! Flexible and sparse representations of mean-field interactions Grid-free approximation methods for first-order problems Compatibility with powerful convex optimization techniques such as ADMM, PDHG the system It is !39

Slide 40

Slide 40 text

Examples: Take a periodic domain ⌦ = T2, and periodic Gaussian interaction kernels K ,µ (x, y) = 2µ2 ⇡ 2 X ↵2Z2 e 2|x y ↵|2 2 , x, y 2 T2. (17) Furthermore, suppose that H(x, p) = |p|2 2 , (x, p) 2 T2 ⇥ R2, D = 0. (18) Observation: Trigonometric system yields diagonal (sparse) representations for translation invariant kernels! !40

Slide 41

Slide 41 text

Figure: Gaussian kernels for ( , µ) = (0.1, 0.75), (0.1, 0.5), (1, 0.5). Figure: The terminal distribution ⇢T (x1, x2 ) and the initial cost 0 (x1, x2 ). !41

Slide 42

Slide 42 text

!42

Slide 43

Slide 43 text

!43

Slide 44

Slide 44 text

!44

Slide 45

Slide 45 text

Numerical method 2D !45

Slide 46

Slide 46 text

!46

Slide 47

Slide 47 text

!47

Slide 48

Slide 48 text

Constrained MFG Given a parameterized density ⇢(✓, x) , consider the variational problem @t⇢(✓(t), x) + r · (⇢(✓(t), x)v(t, x)) = 0, ✓(T) fixed such that the constrained continuity equation holds Typical parameterized density are given by probability models, such as mixed Gaussian, deep neural networks, graphical models etc. inf ⇢(✓,·) Z T 0 n Z L(x, v)⇢(✓(t), x)dx F(⇢(✓(t), ·)) o dt + G(⇢(✓(0), ·)) !48

Slide 49

Slide 49 text

submanifold !49

Slide 50

Slide 50 text

Evolutionary dynamics for Robotics min ⇢ KL(⇢k⇢ref ) + 1 2 Z Z W(x, y)⇢(x)⇢(y)dxdy Wasserstein Gradient flow @t⇢ = r · (⇢rW ⇤ ⇢) + r · (⇢r log ⇢ ⇢ref ) Lagrangian coordinates ˙ X = rW ⇤ ⇢ + r log ⇢ ⇢ref !50

Slide 51

Slide 51 text

!51

Slide 52

Slide 52 text

!52

Slide 53

Slide 53 text

Examples !53