290 likes | 488 Views
Linear and Nonlinear Optimization. Ashish Goel Department of Management Science and Engineering Stanford University Stanford, CA 94305, U.S.A . http://www.stanford.edu/~ashishg http://www.stanford.edu/class/msande211 /. 1 st Day Questions.
E N D
Linear and Nonlinear Optimization Ashish Goel Department of Management Science and Engineering Stanford University Stanford, CA 94305, U.S.A. http://www.stanford.edu/~ashishg http://www.stanford.edu/class/msande211/ Lecture #1; Based on slides by Yinyu Ye
1st Day Questions • 6 Homeworks (best five), 1 Midterm (Oct 28, 1 Final) • Regular Students: 35% H + 25% M + 40% F • Extra credit: 5-10%% • Greater than 100% => A+ • No difference in taking 3 or 4 units • No formula for cutoff between A/B etc. • The more fun we all have, the more A’s we will give out. • Required Textbook: Optimization and Algorithms (PDF will be available) • Website: http://web.stanford.edu/class/msande211 • Programming ≠ Computer Science. The software use will help: Mostly an Excel and a “PAPER AND PENCIL” class! Extra credit Matlab assignments. • Friday’s problem sessions • CA team Lecture #1; Based on slides by Yinyu Ye
Introduction to Optimization • Often considered the common goal of management science & engineering. • Maximize or Minimize f(x) for all x ϵ some set X • Applications in: • Applied Science, Engineering, Economics, Finance, Medicine, Statistics, Business • General Decision and Policy Making • The famous Eighteenth Century Swiss mathematician and physicist Leonhard Euler (1707-1783) proclaimed that “…nothing at all takes place in the Universe in which some rule of maximum or minimum does not appear.” Lecture #1; Based on slides by Yinyu Ye
The Prototypical Optimization Problem Minimize: f(x) Subject to: h1(x) = 0 ... hm(x) = 0 g1(x) < 0 ... gr(x) < 0 The Function could be: x1+2x2, x2+2xy+2y2, xln(x)+ey, |x|+max{x,y}, etc Lecture #1; Based on slides by Yinyu Ye
An Example: Maximum Flow How much flow can travel from A to B, given that each of the directed connecting routes have flow limits? 12 13 6 5 A B 5 11 4 10 Lecture #1; Based on slides by Yinyu Ye
Maximum Flow by Inspection 11 12 2 4 13 6 15 15 5 1 6 11 4 3 5 10 9 Lecture #1; Based on slides by Yinyu Ye
An Example: Maximum Flow Cut value from Source site to Sink site=17 12 13 6 5 A B 5 11 4 10 Lecture #1; Based on slides by Yinyu Ye
An Example: Maximum Flow Cut value from Source site to Sink site=17 12 13 6 5 A B 5 11 4 10 Lecture #1; Based on slides by Yinyu Ye
An Example: Maximum Flow Cut value from Source site to Sink site=15 12 13 6 5 A B 5 11 4 10 Lecture #1; Based on slides by Yinyu Ye
outflow • inflow 12 2 4 • x12 13 6 • (k12) 5 1 6 5 • x13 • (k13) 11 4 3 5 10 • x61 • (k61) ∞ Lecture #1; Based on slides by Yinyu Ye
3 Main Categories in Optimization • Linear Optimization (Programming) • Search Algorithms • Interior-Point Algorithms • Unconstrained Optimization • Constrained Optimization Other Classifications: Quadratic, Convex, Integer, Mixed-Integer, Binary, etc. Lecture #1; Based on slides by Yinyu Ye
Issues in Optimization • Problem Size • Small – by hand • Medium – by software • Large –by decomposition • Algorithm Complexity • Convergence • Local Convergence Time • Insight more than just the solution? • Solution structure properties • Sensitivity analysis • Alternate formulations Lecture #1; Based on slides by Yinyu Ye
What do you learn? • Models – the art • How we choose to represent real problems • Theory – the science • Necessary and Sufficient Conditions that must be true for the optimality of different classes of problems. • Algorithms – the tools • How we apply the theory to robustly and efficiently solve problems and gain insight beyond the solution. Lecture #1; Based on slides by Yinyu Ye
Section I: Linear Programming • Why do we study LP’s • Not because solving non-linear problems are too difficult • But because real-world problems are often formulated as linear equations • Either because they indeed are linear • Or because it is unclear how to represent them and linear is an intuitive compromise Lecture #1; Based on slides by Yinyu Ye
LP, Nobel Prize,… Lecture #1; Based on slides by Yinyu Ye
… and National Medal of Science Lecture #1; Based on slides by Yinyu Ye
Art of Modeling & Vocabulary • Decision Variables x ϵ Rn, yet to be decide • Coefficients, c ϵ Rn, that are given and fixed • Objective inner product f =cTx: Rn ⇥R • Constraint Set S ⊂ Rn • Feasible solution x ϵ S • Optimal solution x* ϵ S • Optimal value z* = f (x*) Lecture #1; Based on slides by Yinyu Ye
x2 • x1 • x1 • x1 LP Example 1: Production Management Lecture #1; Based on slides by Yinyu Ye
Abstract Model • x11 • 1 • 1 • 2 • 2 • 3 • 3 • x34 • 4 LP Example 2: Transportation Lecture #1; Based on slides by Yinyu Ye
LP Example 3: Support Vector Machine • ai • bj {y: yTx + x0 = 0} x is the normal direction or slope vector and x0 is the intersect Lecture #1; Based on slides by Yinyu Ye
LP Example 3: Is Strict Separation Possible Are there x and x0 such that the following (open) inequalities are all satisfied Are there x and x0 such that the following inequalities are all satisfied for arbitrarily small ε. Divide x and x0 by ε., the problem can equivalently reformulated. This is a special LP, called linear feasibility problem. Lecture #1; Based on slides by Yinyu Ye
LP Example 4: Electric Vehicle Charging Schedule Lecture #1; Based on slides by Yinyu Ye
LP Example 4: When Discharge is Allowed Lecture #1; Based on slides by Yinyu Ye
Linear Programming Abstraction Lecture #1; Based on slides by Yinyu Ye
Abstract Linear Programming Model Lecture #1; Based on slides by Yinyu Ye
Coefficient matrix • Obj. vector decision vector • RHS vector LP in Compact Matrix Form Lecture #1; Based on slides by Yinyu Ye
Some Facts of Linear Programming • Adding a constant to the objective function does not change the optimality • Scaling the objective coefficients does not change the optimality • Scaling the right-hand-side coefficients does not change the optimality but the solution gets scaled accordingly • Reordering the decision variables (together with their corresponding objective and constraint coefficients) does not change the optimality • Reordering the constraints (together with their right-hand-side coefficients) does not change the optimality • Multiplying both sides of an equality constraint by a constant does not change the optimality • Pre-multiplying both sides of all equality constraints by a non-singular matrix does not change the optimality Lecture #1; Based on slides by Yinyu Ye
Homework • Before next class read Chapters 1, 2, and 3 in Optimization and Algorithms Lecture #1; Based on slides by Yinyu Ye