Subgradient methods for huge-scale optimization problems

Transcription

1 Subgradient methods for huge-scale optimization problems Yurii Nesterov, CORE/INMA (UCL) May 24, 2012 (Edinburgh, Scotland) Yu. Nesterov Subgradient methods for huge-scale problems 1/24

2 Outline 1 Problems sizes 2 Sparse Optimization problems 3 Sparse updates for linear operators 4 Fast updates in computational trees 5 Simple subgradient methods 6 Application examples 7 Computational experiments: Google problem Yu. Nesterov Subgradient methods for huge-scale problems 2/24

3 Nonlinear Optimization: problems sizes Class Operations Dimension Iter.Cost Memory Small-size All n 4 n 3 Kilobyte: 10 3 Medium-size A n 3 n 2 Megabyte: 10 6 Large-scale Ax n 2 n Gigabyte: 10 9 Huge-scale x + y n log n Terabyte: Sources of Huge-Scale problems Internet (New) Telecommunications (New) Finite-element schemes (Old) PDE, Weather prediction (Old) Main hope: Sparsity. Yu. Nesterov Subgradient methods for huge-scale problems 3/24

4 Sparse problems Problem: min f (x), x Q where Q is closed and convex in RN, and f (x) = Ψ(Ax), where Ψ is a simple convex function: Ψ(y 1 ) Ψ(y 2 ) + Ψ (y 2 ), y 1 y 2, y 1, y 2 R M, A : R N R M is a sparse matrix. Let p(x) def = # of nonzeros in x. Sparsity coefficient: γ(a) def = p(a) MN. Example 1: Matrix-vector multiplication Computation of vector Ax needs p(a) operations. Initial complexity MN is reduced in γ(a) times. Yu. Nesterov Subgradient methods for huge-scale problems 4/24

5 Example: Gradient Method x 0 Q, x k+1 = π Q (x k hf (x k )), k 0. Main computational expenses Projection of simple set Q needs O(N) operations. Displacement x k x k hf (x k ) needs O(N) operations. f (x) = A T Ψ (Ax). If Ψ is simple, then the main efforts are spent for two matrix-vector multiplications: 2p(A). Conclusion: As compared with full matrices, we accelerate in γ(a) times. Note: For Large- and Huge-scale problems, we often have γ(a) Can we get more? Yu. Nesterov Subgradient methods for huge-scale problems 5/24

6 Sparse updating strategy Main idea def After update x + = x + d we have y + = Ax + = }{{} Ax +Ad. y What happens if d is sparse? Denote σ(d) = {j : d (j) 0}. Then y + = y + d (j) Ae j. Its complexity, κ A (d) = M κ A (d) def = j σ(d) j σ(d) p(ae j ), γ(ae j ) = γ(d) 1 p(d) γ(d) max γ(ae j) MN. 1 j m If γ(d) cγ(a), γ(a j ) cγ(a), then κ A (d) c 2 γ 2 (A) MN. j σ(d) can be VERY small! γ(ae j ) MN j σ(d) Expected acceleration: (10 6 ) 2 = sec years! Yu. Nesterov Subgradient methods for huge-scale problems 6/24

7 When it can work? Simple methods: No full-vector operations! (Is it possible?) Simple problems: Let us try: Functions with sparse gradients. 1 Quadratic function f (x) = 1 2 Ax, x b, x. The gradient f (x) = Ax b, x R N, is not sparse even if A is sparse. 2 Piece-wise linear function g(x) = max [ a i, x b (i) ]. Its 1 i m subgradient f (x) = a i(x), i(x) : f (x) = a i(x), x b (i(x)), can be sparse is a i is sparse! But: We need a fast procedure for updating max-type operations. Yu. Nesterov Subgradient methods for huge-scale problems 7/24

8 Fast updates in short computational trees Def: Function f (x), x R n, is short-tree representable, if it can be computed by a short binary tree with the height ln n. Let n = 2 k and the tree has k + 1 levels: v 0,i = x (i), i = 1,..., n. Size of the next level halves the size of the previous one: v i+1,j = ψ i+1,j (v i,2j 1, v i,2j ), j = 1,..., 2 k i 1, i = 0,..., k 1, where ψ i,j are some bivariate functions. v k 1,1 v k,1 v k 1, v 2,1 v 2,n/4 v 1,1 v 1,2... v 1,n/2 1 v 1,n/2 v 0,1 v 0,2 v 0,3 v 0,4 v 0,n 3 v 0,n 2 v 0,n 1 v 0,n Yu. Nesterov Subgradient methods for huge-scale problems 8/24

9 Main advantages Important examples (symmetric functions) f (x) = x ( p, p 1, ψ i,j (t 1, t 2 ) [ t 1 p + t 2 p ] 1/p, n ) f (x) = ln e x(i), ψ i,j (t 1, t 2 ) ln (e t 1 + e t 2 ), i=1 f (x) = max x (i), ψ i,j (t 1, t 2 ) max {t 1, t 2 }. 1 i n The binary tree requires only n 1 auxiliary cells. Its value needs n 1 applications of ψ i,j (, ) ( operations). If x + differs from x in one entry only, then for re-computing f (x + ) we need only k log 2 n operations. Thus, we can have pure subgradient minimization schemes with Sublinear Iteration Cost. Yu. Nesterov Subgradient methods for huge-scale problems 9/24

10 Simple subgradient methods I. Problem: f def = min x Q f (x), where Q is a closed and convex and f (x) L(f ), x Q, the optimal value f is known. Consider the following optimization scheme (B.Polyak, 1967): ( x 0 Q, x k+1 = π Q x k f (x k) f ) f (x k ) 2 f (x k ), k 0. Denote f k = min 0 i k f (x i). Then for any k 0 we have: f k f L(f ) x 0 π X (x 0 ) (k+1) 1/2, x k x x 0 x, x X. Yu. Nesterov Subgradient methods for huge-scale problems 10/24

11 Proof: Let us fix x X. Denote r k (x ) = x k x. r 2 k+1 (x ) Then x k f (x k) f f (x f (x k ) 2 k ) x 2 = rk 2(x ) 2 f (x k) f f (x f (x k ) 2 k ), x k x + (f (x k) f ) 2 f (x k ) 2 r 2 k (x ) (f (x k) f ) 2 f (x k ) 2 r 2 k (x ) (f k f ) 2 L 2 (f ). From this reasoning, x k+1 x 2 x k x 2, x X. Corollary: Assume X has recession direction d. Then x k π X (x 0 ) x 0 π X (x 0 ), d, x k d, x 0. (Proof: consider x = π X (x 0 ) + αd, α 0.) Yu. Nesterov Subgradient methods for huge-scale problems 11/24

12 Constrained minimization (N.Shor (1964) & B.Polyak) II. Problem: min{f (x) : g(x) 0}, where x Q Q is closed and convex, f, g have uniformly bounded subgradients. Consider the following method. It has step-size parameter h > 0. ( ) If g(x k ) > h g (x k ), then (A): x k+1 = π Q x k g(x k) g (x g (x k ) 2 k, ) else (B): x k+1 = π Q (x k h f (x k ) f (x k ). Let F k {0,..., k} be the set (B)-iterations, and = min f (x i ). i F k f k Theorem: If k > x 0 x 2 /h 2, then F k and f k f (x) hl(f ), max g(x i ) hl(g). i F k Yu. Nesterov Subgradient methods for huge-scale problems 12/24

13 Computational strategies 1. Constants L(f ), L(g) are known (e.g. Linear Programming) ɛ max{l(f ),L(g)}. We can take h = number of steps N (easy!). Note: The standard advice is h = Then we need to decide on the R N+1 2. Constants L(f ), L(g) are not known Start from a guess. (much more difficult!) Restart from scratch each time we see the guess is wrong. The guess is doubled after restart. 3. Tracking the record value f k Double run. Other ideas are welcome! Yu. Nesterov Subgradient methods for huge-scale problems 13/24

14 Random sparse coordinate methods III. Problem: { } min f (x) def = max [l i(x) a i, x b i ]. x 0 1 i M Define i(x) : f (x) = l i(x) (x), and random variable ξ(x), which gives indexes from σ(a i(x) ) with equal probabilities 1 p(a i(x) ). Assuming that f is known, we can define now a random vector variable Next(x) by the following rules: 1. Compute h(x) = f (x) f. Generate j(x) = ξ(x). f (x) 2 ( ) 2. Define [ Next(x) ] (j(x)) = x (j(x)) h(x)a (j(x)) i(x) 3. For other indexes j j(x), define [ Next(x) ] (j) = x (j). +. Yu. Nesterov Subgradient methods for huge-scale problems 14/24

15 Algorithmic scheme 0. Choose x 0 0. Compute u 0 = Ax 0 b and f (x 0 ). 1. kth iteration (k 0). a) Generate j k = ξ(x k ) and update x k+1 = Next(x k ). b) Update u k+1 = u k + Ae jk parallel the value of f (x k+1 ). ( x (j k) k+1 x (j k) k ), re-computing in This method defines a sequence of discrete random variables {x k }. Denote fk = min f (x i). 0 i k Theorem: Let p(a i ) r, i = 1,..., m. Then, for any k 0 we have: ( E [fk f ] 2) rl2 (f ) x 0 π X (x 0 ) 2 k+1, E( x k x 2 ) x 0 x 2, x X. NB: One iteration needs at most max 1 j N p(ae j) log 2 M operations. Yu. Nesterov Subgradient methods for huge-scale problems 15/24

16 Application examples Observations: 1 Very often, Large- and Huge- scale problems have repetitive sparsity patterns and/or limited connectivity. Social networks. Mobile phone networks. Truss topology design (local bars). Finite elements models (2D: four neighbors, 3D: six neighbors). 2 For p-diagonal matrices κ(a) p 2. Yu. Nesterov Subgradient methods for huge-scale problems 16/24

17 Google problem Goal: Rank the agents in the society by their social weights. Unknown: x i 0 - social influence of agent i = 1,..., N. Known: σ i - set of friends of agent i. Hypothesis Agent i shares his support among all friends by equal parts. The influence of agent i is equal to the total support obtained from his friends. Yu. Nesterov Subgradient methods for huge-scale problems 17/24

18 Mathematical formulation: quadratic problem Let E R N N be an incidence matrix of the connections graph. Denote e = (1,..., 1) T R N and Ē = E diag (E T e) 1. Since, Ē T e = e, this matrix is stochastic. Problem: Find x 0 : Ēx = x, x 0. The size is very big! Known technique: Regularization + Fixed Point (Google Founders, B.Polyak & coauthors, etc.) N09: Solve it by random CD-method as applied to 1 2 Ēx x 2 + γ 2 [ e, x 1]2, γ > 0. Main drawback: No interpretation for the objective function! Yu. Nesterov Subgradient methods for huge-scale problems 18/24

19 Nonsmooth formulation of Google Problem Main property of spectral radius (A 0) If A R+ n n, then ρ(a) = min max 1 e x 0 1 i n x (i) i, Ax. The minimum is attained at the corresponding eigenvector. Since ρ(ē) = 1, our problem is as follows: f (x) def = max 1 i N [ e i, Ēx x (i) ] min x 0. Interpretation: Increase self-confidence! Since f = 0, we can apply Polyak s method with sparse updates. Additional features; the optimal set X is a convex cone. If x 0 = e, then the whole sequence is separated from zero: x, e x, x k x 1 x k = x, e x k. Goal: Find x 0 such that x 1 and f ( x) ɛ. (First condition is satisfied automatically.) Yu. Nesterov Subgradient methods for huge-scale problems 19/24

20 Computational experiments: Iteration Cost We compare Polyak s GM with sparse update (GM s ) with the standard one (GM). Setup: Each agent has exactly p random friends. Thus, κ(a) def = max κ A(A T e i ) p 2. 1 i M Iteration Cost: GM s κ(a) log 2 N p 2 log 2 N, GM pn. (log = 10, log = 20, log = 30) Time for 10 4 iterations (p = 32) N κ(a) GM s GM Time for 10 3 iterations (p = 16) N κ(a) GM s GM sec 100 min! Yu. Nesterov Subgradient methods for huge-scale problems 20/24

21 Convergence of GM s : Medium Size Let N = , p = 16, κ(a) = 576, and L(f ) = Iterations f f Time (sec) Dimension and accuracy are sufficiently high, but the time is still reasonable. Yu. Nesterov Subgradient methods for huge-scale problems 21/24

22 Convergence of GM s : Large Scale Let N = , p = 8, κ(a) = 192, and L(f ) = Iterations f f Time (sec) Final point x : x = , R 2 0 def = x e 2 2 = Theoretical bound: L2 (f )R 2 0 ɛ 2 = Time for GM: 1 year! Yu. Nesterov Subgradient methods for huge-scale problems 22/24

23 Conclusion 1 Sparse GM is an efficient and reliable method for solving Large- and Huge- Scale problems with uniform sparsity. 2 We can treat also dense rows. Assume that inequality a, x b is dense. It is equivalent to the following system: y (1) = a (1) x (1), y (j) = y (j 1) + a (j) x (j), j = 2,..., n, y (n) b. We need new variables y (j) for all nonzero coefficients of a. Introduce p(a) additional variables and p(a) additional equality constraints. (No problem!) Hidden drawback: the above equalities are satisfied with errors. May be it is not too bad? 3 Similar technique can be applied to dense columns. Yu. Nesterov Subgradient methods for huge-scale problems 23/24

24 Theoretical consequences Assume that κ(a) γ 2 (A)n 2. Compare three methods: Sparse updates (SU). Complexity γ 2 (A)n 2 L2 R 2 log n ɛ 2 operations. Smoothing technique (ST). Complexity γ(a)n 2 LR ɛ Polynomial-time methods (PT). Complexity (γ(a)n + n 3 )n ln LR ɛ operations. There are three possibilities. Low accuracy: γ(a) LR ɛ < 1. Then we choose SU. Moderate accuracy: 1 < γ(a) LR ɛ < n 2. We choose ST. High accuracy: γ(a) LR ɛ > n 2. We choose PT. NB: For Huge-Scale problems usually γ(a) 1 n. operations. Yu. Nesterov Subgradient methods for huge-scale problems 24/24