Random coordinate descent algorithms for. huge-scale optimization problems. Ion Necoara

Transcription

1 Random coordinate descent algorithms for huge-scale optimization problems Ion Necoara Automatic Control and Systems Engineering Depart. 1

2 Acknowledgement Collaboration with Y. Nesterov, F. Glineur ( Univ. Catholique Louvain) A. Patrascu, D. Clipici (Univ. Politehnica Bucharest) Papers can be found at: 2

3 Outline Motivation Problem formulation Previous work Random coordinate descent alg. for smooth convex problems Random coordinate descent alg. for composite convex problems Random coordinate descent alg. for composite nonconvex problems Conclusions 3

4 Motivation Recent Big Data applications: (a) internet (e.g. PageRank) (b) support vector machine (c) truss topology design (d) distributed control...gave birth to many huge-scale optimization problems (dimension of variables n ) BUT matrices defining the optimization problem are very sparse! 4

5 Motivation PageRank problem (Google ranking, network control, data analysis) Let E R n n be adjacency matrix (column stochastic, sparse matrix) Find maximal unitary eigenvector satisfying Ex = x Number of variables (pages) n Standard technique: power method calculations of PageRank on supercomputers take about one week! Formulation as an optimization problem: 1 min x R n 2 Ex x 2 s.t. e T x = 1, x 0. E has at most p << n nonzeros on each row 1 min x R n 2 xt Z T Zx+q T x s.t. a T x = b, l x u ( Z sparse!) 5

6 Motivation Linear SVM problem Let zi R m i = 1,...,n be a set of training data points, m << n Two classes of data points zi Find hyperplane a T y = b which separates data points z i in two classes Formulation as optimization problem: min a R m,b R 1 2 a 2 +Ce T ξ s.t. α i (a T z i b) 1 ξ i, ξ i 0 i = 1,...,n α i { 1, 1} the id (label) of the class corresponding to z i n very big (many constraints) 6

7 Motivation The dual formulation for linear SVM: 1 min x R n 2 xt (Z T Z)x e T x s.t. α T x = 0, 0 x Ce. Z R m n,m << n depends on training points z i (columns of Z are α i z i ) or Z R n n with sparse columns Primal solution is recovered via: a = i α ix i z i & b = i (at z i α i )/n 1 min x R n 2 xt Z T Zx+q T x s.t. a T x = b, l x u ( Z sparse!) 7

8 Motivation State-of-the-art: 1. Second-order algorithms (Newton method, Interior point method): solve at least one linear system per iteration Second-order methods Complexity per iteration O(n 3 ) Worst-case no. of iterations O(lnln 1 ǫ )/O(ln 1 ǫ ) where ǫ is the desired accuracy for solving the optimization problem Let n = 10 8, a standard computer with 2GHz processor takes: 10 7 years to finish only 1 iteration! 8

9 Motivation 2. First-order algorithms (Gradient method, Fast-Gradient method) perform at least one matrix-vector multiplication per iteration (in quadratic case) First-order methods Complexity per iteration O(n 2 ) Worst-case no. of iterations O( 1 ǫ )/O( 1 ǫ ) For n = 10 8, a standard computer with 2GHz processor takes hours per iteration and 100 days to attain ǫ = 0.01 accuracy! Conclusion: for n we require algorithms with low complexity per iteration O(n) or even O(1)! Coordinate Descent Methods 9

10 Problem formulation F = min F(x) (= f(x)+h(x)) x Rn Define decompositions: n = N i=1 x = N i=1 s.t. a T x = b (or even Ax = b) coupling constraints n i,i n = [E 1...E N ] with n E i x i R N and x ij = E i x i +E j x j, x i R n i (i) (ii) A R m n, m << n f has block-component Lipschitz continuous gradient, i.e. (iii) i f(x+e i s i ) i f(x) L i s i x R n,s i R n i, i = 1,...,N h nonsmooth, convex and componentwise separable, i.e. n h(x) = h i (x i ) e.g. : h = 0 or h = 1 [l,u] or h = µ x 1... i=1 10

11 Previous work - Greedy algorithms Tseng (2009) developed coordinate gradient descent methods with greedy strategy min F(x) (= f(x)+h(x)) x Rn s.t. a T x = b (or Ax = b) Let J {1,...,N} be set of indices at current iteration x, then define direction: d H (x;j) = arg min s R nf(x)+ f(x),s Hs,s +h(x+s) s.t. a T s = 0, s j = 0 j / J, (1) where H R n n is a positive definite matrix chosen at initial step of algorithm Tseng & Yun, A Block-Coordinate Gradient Descent Method for Linearly Constrained Nonsmooth Separable Optimization, J. Opt. Theory Applications,

12 Previous work - Greedy algorithms Algorithm (CGD): 1. Choose set of indices J k {1,...,N} w.r.t. Gauss-Southwell rule 2. Solve (1) with x = x k, J = J k, H = H k to obtain d k = d Hk (x k ;J k ) 3. Choose stepsize α k > 0 and set x k+1 = x k +α k d k Procedure of choosing J k (Gauss-Southwell rule): (i) (ii) (iii) decompose projected gradient direction d k into low-dimensional vectors evaluate function (1) in each low-dim. vector choose the vector with smallest evaluation and assign to J its support Alg. (CGD) takes O(n) operations per iteration (for quadratic case & A = a) An estimate for rate of convergence of objective function values is: ( nl x 0 x 2 ) O, L = maxl i ǫ i Recently Beck (2012) developed a greedy coordinate descent algorithm (approx. same complexity) for singly linear constrained models with h box indicator function 12

13 Previous work - Random algorithms Nesterov (2010) derived complexity estimates of random coordinate descent methods min f(x) x Q Q = Q 1 Q N convex h(x) = 1 Q (x) f convex and block-component Lipschitz gradient a = 0 (no coupling constraints) Algorithm (RCGD): 1. Choose randomly and index i k with respect to given probability p ik 2. Set x k+1 = x k +E ik ik f(x k ). We can choose Lipschitz dependent probabilitiesp i = L i / N i=1 L i For structured cases (sparse matrices with p << n number of nonzeros per row) has complexity per iteration O(p)! 13

14 Previous work - Random algorithms An estimate for rate of convergence for the expected values of objective function for Nesterov s method (RCGD) is O ( N i=1 L ) i x 0 x 2 Richtarik (2012), Lu (2012) extended complexity estimates of Nesterov s random coordinate descent method to composite case ǫ min F(x) (= f(x)+h(x)) x Rn f convex and has block-component Lipschitz gradient h nonsmooth, convex, block-separable parallel implementations & inexact implementations were also analyzed Y. Nesterov, Efficiency of coordinate descent methods on huge-scale optimization problems, SIAM J. Opt.,

15 Random coordinate descent - smooth & constrained case min x R n f(x) s.t. a T x = b (or Ax = b) f convex & has block-component Lipschitz gradient communication via connected graph G = (V,E) Algorithm (RCD) : given x 0, a T x 0 = b 1. Choose randomly a pair (i k,j k ) E with probability p ik j k 2. Set x k+1 = x k +E ik d ik +E jk d jk, d ij = (d i,d j ) = arg min s ij R n i +n j f(x)+ ij f(x),s ij + L i +L j 2 s ij 2 s.t. a T i s i +a T j s j = 0 each iteration requires approximately O(p) operations (quadratic case)! Necoara, Nesterov & Glineur, A random coordinate descent method on large optimization problems with linear constraints, ICCOPT, 2013 Necoara, Random coordinate descent algorithms for multi-agent convex optimization over networks, IEEE Trans. Automatic Control,

16 (RCD) smooth case - convergence rate Characteristics: only 2 components (in E) of x are updated per iteration (distributed!) alg. (RCD) needs only 2 components of gradient complexity per iterationo(p)! ( ) closed-form solution e.g. di = 1 L i +L i f(x) at ij ijf(x) a i j a T ij a ij Theorem 1 Let x k generated by Algorithm (RCD). Then, the following estimates for expected values of objective function can be obtained E[f(x k )] f x0 x 2 If additionaly, function f is σ-strongly convex, then λ 2 (Q)k E[f(x k )] f ( 1 λ 2 (Q)σ) k(f(x 0 ) f ) where Q = (i,j) E ( ) p ij I L i +L ni +n j j a ija T ij a T ij a ij (Laplacian matrix of the graph) 16

17 Selection of probabilities I. uniform probabilities: p ij = 1 E II. probabilities dependent on the Lipschitz constants L i p α ij = Lα ij L α, where Lα = (i,j) E L α ij, α 0. III. optimal probabilities obtained from max Q M λ 2 (Q) SDP [p ij ] (i,j) E = argmax t, Q } {t : Q+t aat a T a ti n, Q M. M = {Q R n n : Q = (i,j) E p ij L ij Q ij, p ij = p ji, p ij = 0 if (i,j) E, (i,j) E p ij = 1}. 17

18 Comparison with full projected gradient alg. Assume: then a = e and Lipschitz dependent probabilities p 1 ij = L i +L j L 1 Q = 1 n i=1 L i (I n 1n eet ) λ 2 (Q) = 1 n i=1 L i Alg. (RCD) Alg. full projected grad. iter. complexity O(p) iter. complexity O(n p) E[f(x k )] f i L i x 0 x 2 f(x k k ) f L f x 0 x 2 k 2 f(x) L f I n Remark: maximal eigenvalue of a symmetric matrix can reach its trace! worst case: rate of convergence of (RCD) met. is the same as of full gradient met.! However: - (RCD) method has cheap iteration - (RCD) method has more chances to accelerate 18

19 Numerical tests (I) - Google problem Google problem: mine T x=1 Ex x 2 accuracy ǫ = 10 3, full iterations: x 0,x n 2,x n,,x kn * P ij 1 P ij P ij 0 P ij 0 P ij A*x k x k / x k A*x k x k / x k k k Equivalent number of full iterations versus Ex k x k / x k Left: n = 30 using probabilities p 0 ij,p1 ij and p ij Right: n = 10 6 using probabilitiesp 0 ij and p1 ij 19

20 Random coordinate descent - composite case min x R nf(x)+h(x) s.t. a T x = b f convex with block-component Lipsch. gradient h convex, nonsmooth and separable: h(x) = n (e.g. h = 1 [l,u] or h = µ x 1...) i=1 h i (x i ) Algorithm (CRCD): a T x 0 = b 1. Choose randomly a pair (i k,j k ) with probability p ik j k 2. Set x k+1 = x k +E ik d ik +E jk d jk, d ij = (d i,d j ) = arg min s ij R n i +n j f(x)+ ij f(x),s ij + L i +L j 2 s ij 2 +h(x+s ij ) s.t. a T i s i +a T j s j = 0 each iteration requires approximately O(p) operations (quadratic case)! 20

21 Random coordinate descent - composite case Characteristics: only 2 components of x are updated per iteration alg. (CRCD) needs only 2 components of the gradient and is using only 2 functions h i & h j of h if N = n and h is given by l1 norm or indicator function for box, then the direction d ij can be computed in closed form if N < n and h is coordinatewise separable, strictly convex and piece-wise linear/quadratic with O(1) pieces (e.g. h given by l 1 norm), then the direction d ij can be computed in linear-time (i.e. O(n i +n j ) operations). the complexity of choosing randomly a pair (i,j) with a uniform probability distribution requires O(1) operations 21

22 (CRCD) composite case - convergence rate Theorem 2 Let x k be generated by Algorithm (CRCD) and L = max i L i. If the index pairs are selected with uniform distribution, then we have E[F(x k )] F N2 L x 0 x 2. k If additionaly, function f is σ-strongly convex, then E[F(x k )] F ( 1 2(1 γ) ) k N 2 (F(x 0 ) F ), where γ is defined by: γ = 1 σ 2L σ, 8L, if σ 4L otherwise. Necoara & Patrascu, Random coordinate descent algorithm for optimization problems with composite objective function and linear coupled constraints, Computational Opt. Appl.,

23 Arithmetic complexity - comparison N = n (scalar case) & sparse QP R 2 = x x 0 2 L f i L i & L = max i L i & L av = i L i n metoda grad. Lipsc. model complexity per iteration GM (Nesterov) O( L fr 2 ) ǫ h & a full gradient - O(n) CGM (Tseng) O( nlr2 ) ǫ h & a partial gradient - O(n) RCGM (Nesterov) O( nl avr 2 ) ǫ h & a = 0 partial gradient - O(1) RCD O( nl avr 2 ) ǫ h = 0 & a partial gradient - O(1) CRCD O( n2 L av R 2 ) ǫ h & a partial gradient - O(1) our methods RCD & CRCD have usually better (N < n) or comparable (N = n) arithmetic complexity than (or with) existing methods adequate for parallel or distributed architectures robust and have more chances to accelerate (due to randomness) easy to implement (closed-form solution) 23

24 Numerical tests (II) - SVM problem Data set n/m (CRCD) full-iter/obj/time(s) (CGD) iter/obj/time(s) a7a 16100/122 (p = 14) 11242/ / / /21.5 a9a 32561/123 (p = 14) 15355/ / / /89.0 w8a 49749/300 (p = 12) 15380/ / / /27.2 ijcnn /22 (p = 13) 7601/ / / /16.5 web /254 (p = 85) 1428/ / / /748 covtyp /54 (p = 12) 1722/ / /-24000/480 test /10 6 (p = 50) 228/ / / /568 test /10 3 (p = 10) 500/ / / /516,66 real test problems taken from LIBSVM library Our alg. (CRCD) - by a factor of 10 faster than (CGD) method (Tseng)! 24

25 Numerical tests (III) - Chebyshev center problem Chebyshev center problem: given a set of points z 1,...,z n R m, find the center z c and radius r of the smallest enclosing ball of the given points Applications: pattern recognition, protein analysis, mechanical engineering Formulation as an optimization problem: min r,z c r s.t.: z i z c 2 r i = 1,...,n, where r is the radius and z c is the center of the enclosing ball. Dual problem: min x R n Zx 2 s.t. e T x = 1, n z i 2 x i +1 [0, ) (x) (2) i=1 where Z contains the given points z i as columns 25

26 Numerical tests (III) - Chebyshev center problem Simple recovery of primal optimal solution from dual x : r = ( Zx 2 + n z i 2 x i i=1 ) 1/2, z c = Zx. (3) Two sets of numerical experiments: all alg. start from x 0 = e 1 : observe that Tseng s algorithm has good performance and Gradient Method is worst starting from x 0 = e/n: observe that Gradient Method has good performance and Tseng is worst algorithm (CRCD) is very robust w.r.t. starting point x 0 26

27 Numerical tests (III) - Chebyshev center problem x 0 = e/n (CGD) x 0 = e/n (GM) x 0 = e/n (CRCD) x 0 = e 1 x 0 = e 1 x 0 = e 1 27

28 Random coordinate descent - nonconvex & composite min f(x)+h(x) x R n f nonconvex with block-component Lip. gradient & a = 0 (no coupling constraints) h is proper, convex and block separable e.g. : h = 0 or h = 1 [l,u] or h = µ x 1... If x R n is a local minimum, then the following relation holds 0 f(x )+ h(x ) (stationary points) Algorithm (NRCD): 1. Choose randomly an index i k with probability p ik 2. Compute x k+1 = x k +E ik d ik d i = arg min s i R n i f(x)+ i f(x),s i + L i 2 s i 2 +h(x+s i ). Each iteration is cheap, complexity O(p), where p << n (even closed-form solution)! Patrascu & Necoara, Efficient random coordinate descent algorithms for large-scale structured nonconvex optimization, submitted J. Global Opt.,

29 (NRCD) nonconvex & composite - convergence rate We introduce the following map (here L = [L 1...L N ] and D L = diag(l)): d L (x) = arg min s R nf(x)+ f(x),s s 2 L +h(x+s) We define optimality measure: M 1 (x,l) = D L d L (x) L, where s 2 L = st D L s and u L its dual norm (observe that M 1(x,L) = 0 x stationary point) Theorem 3 Let the sequence x k be generated by Algorithm (NRCD) using the uniform distribution, then the following statements are valid: (i) (ii) The sequence of random variables M 1 (x k,l) converges to 0 a.s. and the sequence F(x k ) converges to a random variable F a.s. Any accumulation point of the sequence x k is a stationary point Moreover, in expectation [ ( ] 2 min E M 1 (x,l)) l 2N ( F(x 0 ) F ) 0 l k k k 0 29

30 Random coordinate descent - nonconvex & constrained min x R nf(x)+h(x) s.t. a T x = b, f nonconvex with block-component Lip. gradient h is proper, convex and separable If x is a local minimum, then there exists a scalar λ such that: 0 f(x )+ h(x )+λ a and a T x = b. Algorithm (NCRCD): 1. Choose randomly a pair (i k,j k ) with probability p ik j k 2. Compute x k+1 = x k +E ik d ik +E jk d jk, d ij = (d i,d j ) = arg min s ij R n i +n j f(x)+ ij f(x),s ij + L i +L j 2 s ij 2 +h(x+s ij ) s.t. a T i s i +a T j s j = 0 Each iteration is cheap, complexity O(p) (even closed-form solution)! 30

31 (NCRCD) nonconvex & constrained - convergence rate We introduce the following map: d T(x) = arg min f(x)+ f(x),s + 1 s R n : a T s=0 2 s 2 T +h(x+s). We define the optimality measure: M 2 (x,t) = D T d NT (x) T, where T i = 1 N j L ij (observe that M 2 (x,t) = 0 x stationary point) Theorem 4 Let the sequence x k be generated by Algorithm (NCRCD) using the uniform distribution, then the following statements are valid: (i) (ii) The sequence of random variables M 2 (x k,t) converges to 0 a.s. and the sequence F(x k ) converges to a random variable F a.s. Any accumulation point of the sequence x k is a stationary point Moreover, in expectation [ ( ] 2 min E M 2 (x,t)) l N ( F(x 0 ) F ) 0 l k k k 0. 31

32 Numerical tests (IV) - eigenvalue complementarity problem Eigenvalue complem. prob. (EiCP): given matrices A,B R n n, find λ R & x 0 w = (λb A)x, w 0, x 0, w T x = 0 Applications of EiCP: optimal control, stability analysis of dynamic systems, electrical networks, quantum chemistry, chemical reactions, economics... If A, B are symmetric, then we have symmetric (EiCP). Symmetric (EiCP) is equivalent with finding a stationary point of a generalized Rayleigh quotient on the simplex: x T Ax min x R n x T Bx s.t.: et x = 1, x 0. Equivalent nonconvex logarithmic formulation (for A,B 0, with a ii,b ii > 0 e.g. stability of positive dynamical systems): ( ) max f(x) = lnx T Ax lnx T Bx x R n s.t.: e T x = 1, x 0 h(x) = 1 [0, ) (x) Perron-Frobenius theory for A irreducible and B = I n implies global maximum! 32

33 Numerical tests (IV) - eigenvalue complementarity problem Compare with DC (Difference of Convex functions) algorithm (Thi et al. 2012), equivalent in some sense with Projected Gradient method Hard to estimate Lipschitz parameter µ in DC alg., but crucial for convergence of DC ( µ µ ) max x: e T x=1, x 0 2 x2 +lnx T Ax lnx Bx) ( T 2 x2 DC NCRCD n µ CPU (sec) Iter F CPU (sec) Iter F n n n n n n n n n n n n

34 Conclusions usually full first/second-order methods are inefficient for huge-scale optimization for sparse problems coordinate descent methods are adequate for their low complexity per iteration (O(p)) randomized coordinate descent methods have simple strategy for choosing the working set - O(1) operations for index choice usually randomized methods outperform greedy methods we provide rates of convergence and arithmetic complexities for randomized coordinate descent methods randomized methods are easy to implement and adequate for modern parallel and distributed architectures 34

35 References A. Beck, The 2-coordinate descent method for solving double-sided simplex constrained minimization problems, Technical Report, Y. Nesterov, Efficiency of coordinate descent methods on huge-scale optimization problems, SIAM Journal on Optimization 22(2), , P. Richtarik and M. Takac, Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function, Mathematical Programming, Series A, DOI /s z, P. Tseng and S. Yun, A Coordinate Gradient Descent Method for Nonsmooth Separable Minimization, Mathematical Programming, 117, , P. Tseng and S. Yun, A Block-Coordinate Gradient Descent Method for Linearly Constrained Nonsmooth Separable Optimization, Journal of Optimization Theory and Applications, 140, ,

36 References I. Necoara, Y. Nesterov and F. Glineur, A random coordinate descent method on large optimization problems with linear constraints, Technical Report, University Politehnica Bucharest, 2011, I. Necoara, Random coordinate descent algorithms for multi-agent convex optimization over networks, IEEE Transactions on Automatic Control, 58(7), 1-12, I. Necoara and A. Patrascu, A random coordinate descent algorithm for optimization problems with composite objective function and linear coupled constraints, Computational Optimization and Applications, in press, 2013, A. Patrascu and I. Necoara, Efficient random coordinate descent algorithms for large-scale structured nonconvex optimization, submitted to Journal of Global Optimization,