15-451 Algorithms 11/02/00 * Linear programming Quiz #2 next Thurs. 1 sheet notes. ============================================================================ In last couple of classes we looked at: - Bipartite matching: Given a bipartite graph, find the largest set of edges with no endpoints in common. - Network flow. (More general than bipartite matching) - Min-Cost Max-flow (even more general than plain max flow). Today, we'll look at something even more general: Linear Programming. Here, "programming" is meant like "program of action" as in DP. Linear Programming is important because it is very powerful: many many problems can be coded up as linear programs. Especially problems of allocating resources and a lot of business applications. In GSIA there are entire courses devoted to linear programming. We're only going to have time for 1 lecture. So, will just have time to say what they are, and give examples of encoding problems as LPs, will only get a chance to say a little bit about algorithms for solving them. Before defining, motivate with an example: There are 168 hours in a week. Say we want to allocate our time between classes and studying (S), going to parties (P), and everything else (E) (eating, sleeping, taking showers, etc). To survive need to spend at least 56 hours on E (8 hrs/day). To pass courses, need S >= 60, but more if don't sleep enough or spend too much time partying: 2S+E-3P >= 150. (e.g., if don't party at all then this isn't a problem, but if we spend more time on P then need to sleep more or study more). To maintain sanity need P+E >= 70. Q1: can we do this? Formally, is there a *feasible* solution? A: yes. For instance, one feasible solution: S=80, P=20, E = 68 Q2: What is the maximum partying that we can do? Can we do P=30? No: the 2S+E-3P >=150 constraint becomes 2S+E >= 240. Even if set E=56 (minimum) then need S = 92, which is too many hours. So, somewhere in between. Q3: maybe our notion of happiness is expressed by 2P+E. What is a feasible solution such that this is maximized? The formula "2P+E" is called an *objective function*. This is called a *linear program*. What makes it linear is that all our constraints were linear in our variables. E.g., 2S+E-3P>=150. And our objective function is also linear. Not allowed things like requiring S*E >= 100, since this wouldn't be linear. More formally, here is the definition of the linear programming problem DEFINITION OF LINEAR PROGRAMMING PROBLEM: ======================================== Have n variables x_1,..., x_n. Have m linear inequalities in these variables (equalities OK too): E.g., 3x_1 + 4x_2 <= 6 x_1 >= 0 etc. May also have an objective function: 2x_1 + 3x_2 + x_3. Goal: find values for the x_i's that satisfy the constraints and maximizes the objective. "feasibility problem": no objective, just want to satisfy the constraints. For instance, let's write out our time allocation problem this way. Variables are S, P, E. objective is to maximize 2P+E. subject to these constraints: S + P + E = 168 E >= 56 S >= 60 2S+E-3P >= 150 P+E >= 70 P >= 0 (can't spend negative time partying) MORE EXAMPLES OF MODELLING PROBLEMS AS LINEAR PROGRAMS: ====================================================== Typical Operations-Research kind of problem (from Mike Trick): Have 4 production plants for making cars. Each works a little differently in terms of labor needed, materials, and pollution produced per car: labor materials pollution plant 1 2 3 15 plant 2 3 4 10 plant 3 4 5 9 plant 4 5 6 7 We need to produce 400 cars at plant 3 according to labor agreement. Have 3300 hours of labor, 4000 units of material available. Allowed to produce 12000 units of pollution. Want to maximize number of cars produced. How to model? First step: What are the variables? x_1,x_2,x_3,x_4, where x_i = # cars at plant i. Second step: What is our objective? maximize x_1+x_2+x_3+x_4 Last step: what are the constraints? x_i >= 0 for all i x_3 >= 400 2x_1 + 3x_2 + 4x_3 + 5x_4 <= 3300 3x_1 + 4x_2 + 5x_3 + 6x_4 <= 4000 15x_1 + 10x_2 + 9x_3 + 7x_4 <= 12000 MAX FLOW: Variables: set up one variable x_e per edge e. Let's just represent the positive flow. Objective: maximize SUM x_e edges e into T Constraints: For all e, x_e <= c_e and x_e >= 0. [c_e is capacity of edge e] For each node v except S and T, SUM x_e = SUM x_e' edges e into v edges e' leaving v Example: S-->A, cap = 4. A-->C, cap=3. C-->T, cap = 2. S-->B, cap = 2. B-->D, cap=3. D-->T, cap = 4. C-->B, cap = 1. B-->C, cap=2. LP is: maximize x_ct + x_dt subject to 0 <= x_sa <= 4, etc. x_sa = x_ac, x_sb + x_cb = x_bc + x_bd, x_ac + x_bc = x_cb + x_ct, x_bd = x_dt How about min cost max flow? First solve for max flow f. Then add a constraint that flow must equal f, and minimize linear cost function sum_e cost(e)*flow(e). 2-PLAYER ZERO-SUM GAMES: Here is payoff matrix for row player: 20 -10 5 5 10 -10 -5 0 10 What is minimax optimal strategy for row? Use x_1, x_2, x_3 are probabilities on rows 1, 2, and 3 respectively. Want to maximize the worst case (minimum), over all columns opponent can play, of expected gain. Can do like this: maximize v such that: x's are a prob dist: x_1 + x_2 + x_3 = 1 x_1 >= 0, x_2 >= 0, x_3 >=0 get at least v if col1: 20*x_1 + 5*x_2 -5*x_3 >= v ..... if col2: -10*x_1 + 10*x_2 >= v ..... if col3: 5*x_1 - 10x_2 + 10x_3 >= v How to solve linear programs? History: the standard algorithm for solving LPs the Simplex Algorithm, developed in the 40s. It's *not* guaranteed to run in polynomial time, and you can come up with bad examples for it, but in general it runs pretty fast. Only much later in 1980 was it shown that linear programming could be done in polynomial time by something called the Ellipsoid Algorithm (but it is pretty slow). Later on, a faster polynomial-time algorithm Karmarkar's Alg was developed, which is competitive with Simplex. There are a lot of commercial LP packages, for instance LINDO, CPLEX, Solver (in Excel). Won't have time to describe any of these in detail. Instead, give some intuition by viewing LP as a geometrical problem. Think of an n-dimensional space with one coordinate per variable. A solution is a point in this space. An inequality, like x_1 + x_2 <= 6 is saying that we need solution to be on a specified side of a certain hyperplane. Feasible region is the convex region in space defined by these constraints. Then we want to find the feasible point that is farthest in the "objective" direction. Let's go to first example with S, P, E. To make this easier to draw, can use our first constraint that S+P+E = 168 to replace S with 168-P-E. So, just draw in 2 dimensions: P and E. Constraints are: E >= 56 P+E >= 70 P >= 0 S >= 60 which means 168-P-E >= 60 or P+E <= 108. 2S-3P+E >= 150 which means 2(168-P-E)-3P+E >= 150 or 5P+E <= 186 For objective of max P, this happens at E=56, P = 26. For objective of max 2P+E, this happens at P=19.5, E=88.5 Can use this view to motivate the algorithms. SIMPLEX ALG: Earliest and most common algorithm called Simplex method. Idea: start at some "corner" of the feasible region (to make this easier, we can add in "slack variables" that will drop out when we do our optimization). Then we repeatedly do the following step: look at all neighboring corners and go to the best one if it is better. Stop when we get to a corner where no neighbor is better than we are. Neat fact is that since the objective is linear, the optimal solution will be at a corner (or maybe multiple corners). Furthermore, there are no local maxima: if you're *not* optimal then some neighbor of you is better than you are. That's because this is a convex region. So, Simplex method is guaranteed to halt at the best solution. The problem is that it is possible for there to be an exponential number of corners and it is possible for Simplex to take an exponential number of steps to converge. But, in practice this usually works suboptimal corners that are better than all their neighbors. ELLIPSOID ALGORITHM: Invented by Khachiyan in 1980 in Russia. Solve "feasibility problem" (Then can do binary search with our objective function). Start with big ellipse (called an ellipsoid in higher dimensions) that we can be sure contains the feasible region. Then try the center of the ellipse -- see if it violates any constraints. If not, you're done. If so, look at the constraint violated. So we know the solution (if any) is contained in the remaining half-ellipse. Now, find a new smaller ellipse that contains the half-ellipse remaining. Can show the new smaller ellipse has a volume that's significantly smaller. If ever get to too small volume can prove there can't be a solution. One nice thing about ellipsoid algorithm is you just need to tell if the current solution violates any constraints or not, and if so, to produce one. Don't need to explicitly write them all down. KARMARKAR'S ALGORITHM: Sort of has aspects of both. Works with feasible points but doesn't go from corner to corner. Instead it moves inside the interior of the feasible region. One of first of a whole class of so-called "interior-point methods". Development of better and better algorithms is a big ongoing area of research. In particular, get a lot of mileage by using good data structures to speed up time in making each decision.