Constrained Optimization Rong Jin
Outline Equality constraints Inequality constraints Linear Programming Quadratic Programming
Optimization Under Equality Constraints Maximum Entropy Model English ‘in’ French {dans (1), en (2), à (3), au cours de (4), pendant (5)}
Reducing variables Representing variables using only p 1 and p 4 Objective function is changed Solution: p 1 = 0.2, p 2 = 0.3, p 3 =0.1, p 4 = 0.2, p 5 = 0.2
Maximum Entropy Model for Classification It is unlikely that we can use the previous simple approach to solve such a general Solution: Lagrangian
Equality Constraints: Lagrangian Introduce a Lagrange multiplier for the equality constraint Construct the Lagrangian Necessary condition A optimal solution for the original optimization problem has to be one of the stationary point of the Lagrangian
Example: Introduce a Lagrange multiplier for constraint Construct the Lagrangian Stationary points
Lagrange Multipliers Introducing a Lagrange multiplier for each constraint Construct the Lagrangian for the original optimization problem
Lagrange Multiplier We have more variables p 1, p 2, p 3, p 4, p 5 and, 1, 2, 3 Necessary condition (first order condition) A local/global optimum point for the original constrained optimization problem a stationary point of the corresponding Lagrangian Original Entropy Function Constraints
Stationary Points for Lagrangian All probabilities p 1, p 2, p 3, p 4, p 5 are expressed as functions of Lagrange multipliers s
Dual Problem p 1, p 2, p 3, p 4, p 5 are expressed as functions of s We can even remove the variable 3 However, it is still difficult to obtain the solution s such that the constraints are satisfied Dual problem Substitute the expression for ps into the Lagrangian Find the s that MINIMIZE the substituted Lagrangian
Dual Problem Finding s such that the above objective function is minimized Original Lagrangian Substituted Lagrangian Expression for ps
Dual Problem Using dual problem Constrained optimization unconstrained optimization Need to change maximization to minimization Only valid when the original optimization problem is convex/concave (strong duality) Dual Problem Primal Problem x*= * When convex/concave
Maximum Entropy Model for Classification Introduce a Lagrange multiplier for each linear constraint
Maximum Entropy Model for Classification Construct the Lagrangian for the original optimization problem Original Entropy Function Consistency Constraint Normalization Constraint
Stationary Points Stationary points: first derivatives are zero Sum of conditional probabilities must be one Conditional Exponential Model !
Support Vector Machine Having many inequality constraints Solving the above problem directly could be difficult Many variables: w, b, Unable to use nonlinear kernel function
Inequality Constraints: Modified Lagrangian Introduce a Lagrange multiplier for the inequality constraint Construct the Lagrangian Kuhn-Tucker condition A optimal solution for the original optimization problem will satisfy the following conditions Non-negative Lagrange Multiplier Two cases: 1.g(x) = c, 2.g(x) > c =0
Example: Introduce a Lagrange multiplier for constraint Construct the Lagrangian KT conditions Expressing objective function using Solution is =3
SVM Model Lagrange multipliers for inequality constraints Min Max +
SVM Model Lagrangian for SVM model Kuhn-Tucker condition
Dual Problem for SVM Express w, b, using and Finding solution satisfying KT conditions is difficult
Dual Problem for SVM Rewrite the Lagrangian function using only and Simplify using KT conditions
Dual Problem for SVM Final dual problem Maximize Minimize
Quadratic Programming Find Subject to
Linear Programming Find Subject to Very very useful algorithm papers 100+ books 10+ courses 100s of companies Main methods Simplex method Interior point method Most important: how to convert a general problem into the above standard form
Example Need to change max to min Find Subject to
Example Need change to Find Subject to
Example Need to convert the inequality Find Subject to
Example Need change |x 3 | Find Subject to