COT 6936: Topics in Algorithms! Giri Narasimhan. ECS 254A / EC 2443; Phone: x3748

Size: px

Start display at page:

Download "COT 6936: Topics in Algorithms! Giri Narasimhan. ECS 254A / EC 2443; Phone: x3748"

Kerry Dennis
5 years ago
Views:

1 COT 6936: Topics in Algorithms! Giri Narasimhan ECS 254A / EC 2443; Phone: x3748 giri@cs.fiu.edu

2 Gaussian Elimination! Solving a system of simultaneous equations x 1-2x 3 = 2 x 2 + x 3 = 3 x 1 + x 2 - x 4 = 4 x 2 + 3x 3 + x 4 = 5 O(n 3 ) algorithm x 1-2x 3 = 2 x 2 + x 3 = 3 x 2 + 2x 3 - x 4 = 2 x 2 + 3x 3 + x 4 = 5

3 Linear Programming! Want more than solving simultaneous equations We have an objective function to optimize

4 Chocolate Shop [DPV book]! 2 kinds of chocolate milk [Profit: $1 per box] [Demand: 200] Deluxe [Profit: $6 per box] [Demand: 300] Production capacity: 400 boxes Goal: maximize profit Maximize x 1 + 6x 2 subject to constraints: x x x 1 + x x 1, x 2 0

5 Diet Problem! Food type: Nutrients: F 1,,F m N 1,,N n Min daily requirement of nutrients: c 1,,c n Price per unit of food: b 1,,b m Nutrient N j in food F i : a ij Problem: Supply daily nutrients at minimum cost Min Σ i b i x i Σ i a ij x i c j x i 0 for 1 j n

6 Transportation Problem! Ports or Production Units: P 1,,P m Markets to be shipped to: M 1,,M n Min daily market need: r 1,,r n Port/production capacity: s 1,,s m Cost of transporting to M j from port P i : a ij Problem: Meet market need at minimum transportation cost

7 Workers: b 1,,b n Jobs: g 1,,g m Assignment Problem! Value of assigning person b i to job g j : a ij Problem: Choose job assignment to maximize value

8 Bandwidth Allocation Problem! Need: A B 2 units B C 2 units C A 2 units Connections: Short route Long route Revenue: A B pays $3 per unit B C pays $2 per unit C A pays $4 per unit

9 Bandwidth Allocation Problem! Maximize revenue by allocating bandwidth to connections along two routes without exceeding bandwidth capacities Max 3(x AB +x AB ) + 2(x BC +x BC ) + 4(x AC +x AC ) s.t. x AB + x AB + x BC + x BC 10 x AB + x AB + x AC + x AC 12 x BC + x BC + x AC + x AC 8 x AB + x BC + x AC 6; x AB + x AB 2; x BC + x BC 2 x AB + x BC + x AC 13; x AC + x AC 2 x AB + x BC + x AC 11;& all nonneg constraints

10 Standard LP! Maximize Σc j x j [Objective Function] Subject to Σa ij x j b j [Constraints] and x j 0 [Nonnegativity Constraints] Matrix formulation of LP Maximize c T x Subject to Ax b and x 0

11 Converting to standard form! Min -2x 1 + 3x 2 Subject to x 1 + x 2 = 7 x 1 2x 2 4 x 1 0 Max 2x 1-3x 2 Subject to x 1 + x 2 7 -x 1 - x 2-7 x 1 2x 2 4 x 1 0

12 Converting to standard form! Max 2x 1-3x 2 Subject to x 1 + x 2 7 -x 1 - x 2-7 x 1 2x 2 4 x 1 0 Max 2x 1 3(x 3 - x 4 ) Subject to x 1 + x 3 - x 4 7 -x 1 (x 3 - x 4 ) -7 x 1 2(x 3 - x 4 ) 4 x 1, x 3, x 4 0 x 2 is not constrained to be non-negative

13 Converting to Standard form! Max 2x 1 3x 2 + 3x 3 Subject to x 1 + x 2 x 3 7 -x 1 x 2 + x 3-7 x 1 2x 2 2x 3 4 x 1, x 2, x 3 0

14 Slack Form! Max 2x 1 3x 2 + 3x 3 Subject to x 1 + x 2 x 3 7 -x 1 x 2 + x 3-7 x 1 2x 2 2x 3 4 x 1, x 2, x 3 0 Max 2x 1 3x 2 + 3x 3 Subject to x 1 + x 2 x 3 + x 4 = 7 -x 1 x 2 + x 3 + x 5 = -7 x 1 2x 2 2x 3 + x 6 = 4 x 1, x 2, x 3, x 4, x 5, x 6 0

15 Duality! Max c T x Subject to Ax b and x 0 [Primal] Min y T b Subject to y T A c and y 0 [Dual]

x 2 0 (100,300) is feasible; value = 1900. Optimum?

16 Understanding Duality! Maximize x 1 + 6x 2 subject to constraints: x (1) x (2) x 1 + x (3) x 1, x 2 0 (100,300) is feasible; value = Optimum? Adding 1 times (1) + 6 times (2) gives us x 1 + 6x How were mutipliers determined? Adding 1 times (3) + 5 times (2) gives us x 1 + 6x Certificate of Optimality for solution (100,300)

17 Understanding Duality! Maximize x 1 + 6x 2 subject to: x (y 1 ) x (y 2 ) [(100,300)] x 1 + x (y 3 ) x 1, x 2 0 Different choice of multipliers gives us different bounds. We want smallest bound. Minimize 200y y y 3 subject to: y 1 + y 3 1 (x 1 ) y 2 + y 3 6 (x 2 ) [(0,5,1)] y 1, y 2 0

18 Duality Principle! Primal feasible values dual feasible values Max primal value = min dual value Duality Theorem: If a linear program has a bounded optimal value then so does its dual and the two optimal values are equal.

19 Shortest Path Problem as a LP! Graph G = (V,E); Vertices: v 1,,v n ; Edges: e 1,,e m ; Weight function on edges w(e i ); Source s; Dest t; LP: min w T x s.t. A x = b and x 0 Here A and b are defined as follows: A ij = +1 if e j leaves v i ; b s = +1 = -1 if e j enters v i ; b t = -1 = 0 otherwise; b i = 0 else; We want integral solutions for x

20 LP: min w T x Dual LP! s.t. A x = b and x 0 Dual LP: max y s -y t s.t. y u - y v w(e) for every edge e = (u,v)

21 Visualizing Duality! Shortest Path Problem Build a physical model and between each pair of vertices attach a string of appropriate length To find shortest path from s to t, hold the two vertices and pull them apart as much as possible without breaking the strings This is exactly what a dual LP solves! Max x s -x t subject to x u -x v w uv for every edge (u.v) The taut strings correspond to the shortest path, i.e., they have no slack

22 Linear Constraints: Geometric View! 2,3,7 1,3,7 i.e., some inequalities satisfied as equalities Vertex: point where n hyperplanes meet; Neighbor: vertices sharing n-1 hyperplanes

23 Simplex Algorithm! Start at v, any vertex of feasible region while (there is neighbor v of v with better objective value) do set v = v Report v as optimal point and its value as optimal value What is a Vertex?, neighbor? Start vertex? How to pick next neighbor?

24 Steps of Simplex Algorithm! In order to find next neighbor from arbitrary vertex, we do a change of origin (pivot)

25 Simplex Algorithm Example!

26 Simplex Algorithm Example!

27 Simplex Algorithm Example!

28 Simplex Algorithm: Degenerate vertices! i.e., some inequalities satisfied as equalities 2,3,4,5 2,3,7 1,3,7 Vertex: point where n hyperplanes meet; Neighbor: vertices sharing n-1 hyperplanes

29 Polynomial-time algorithms for LP! Simplex is not poly-time in the worst-case Khachiyan s ellipsoid algorithm: LP is in P Karmarkar s interior-point algorithm Good implementations for LP exist Works very well in practice More competitive than the poly-time methods for LP

30 Integer Linear Programming! LP with integral solutions NP-hard If A is a totally unimodular matrix, then the LP solution is always integral. A TUM is a matrix for which every nonsingular submatrix has determinant 0, +1 or -1. A TUM is a matrix for which every nonsingular submatrix has integral inverse.

31 Vertex Cover as an LP?! For vertex v, create variable x v For edge (u,v), create constraint x u + x v 1 Objective function: Σx v Additional constraints: x v 1 Doesn t work because x v needs to be from {0,1}

32 Integer Linear Programming! LP with integral solutions NP-hard If A is a totally unimodular matrix, then the LP solution is always integral. A TUM is a matrix for which every nonsingular submatrix has determinant 0, +1 or -1. A TUM is a matrix for which every nonsingular submatrix has integral inverse.

33 Vertex Cover as an LP?! For vertex v, create variable x v For edge (u,v), create constraint x u + x v 1 Objective function: Σx v Additional constraints: x v 1 Doesn t work because x v needs to be from {0,1}

34 Set Cover! Given a universe of items U = {e 1,, e n } and a collection of subsets S = {S 1,, S m } such that each S i is contained in U Find the minimum set of subsets from S that will cover all items in U (i.e., the union of these subsets must equal U) Weighted Set Cover: Given universe U and collection S, and a cost c(s i ) for each subset S i in S, find the minimum cost set cover

36 Fractional LP may have higher objective U = {e, f, g} S 1 = {e, f} S 2 = {f, g} S 3 = {e, g} function value than integer LP! Optimal set cover = {S 1, S 2 } Fractional optimal set cover assigns ½ to each of three sets giving a total optimal value of 3/2.

38 K-Approximation Alg using Dual Fitting! Ratio to be bounded by K 0 Value of Greedy solution Value of Dual Feasible solution obtained from greedy solution Value of Optimal ILP solution Value of Primal Optimal LP solution = Value of Dual Optimal LP solution

41 Rounding Algorithm for Set Cover! Algorithm Find an optimal solution to the LP Relaxation Pick all sets S for which x S 1/f in this solution f = frequency of most frequent item Analysis Is the resulting solution a valid set cover? How good is the solution? How close is to the optimal set cover?

42 Analysis of Rounding Algorithm!

Motivating examples Introduction to algorithms Simplex algorithm. On a particular example General algorithm. Duality An application to game theory

Instructor: Shengyu Zhang 1 LP Motivating examples Introduction to algorithms Simplex algorithm On a particular example General algorithm Duality An application to game theory 2 Example 1: profit maximization