Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming ISEN 636: Large-Scale Stochastic Optimization Topic: Introduction *** L. Ntaimo Industrial & Systems Engineering Texas A&M University Sept 4, 2009 1 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Outline Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming 2 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Preliminaries I Some Basic Probability First I I I ω is an “outcome" of a random experiment (we will use ω̃ to denote a multivariate random variable vector) Note: The textbook by Birge & Louveaux uses ξ and ξ. Ω is the set of all possible outcomes (sample space) A is collection of random outcomes (events) of Ω 3 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Probability Spaces I For each A ∈ A there is a probability measure (or distribution) P that tells the probability with which A ∈ A occurs I I I I 0 ≤ P(A) ≤ 1 P(Ω) = 1, P(∅) = 0 T P(A1 ∪ A2 ) = P(A1 ) + P(A2 ) if A1 A2 = ∅ The triplet (Ω, A, P) is called a probability space 4 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Random Variables I I I A random variable (rv) ω̃ on a probability space (Ω, A, P) is a real-valued function ω̃(ω), ω ∈ Ω such that {ω|ω̃(ω) ≤ x} is an event for all finite x. For the random variable ω̃ we define its cumulative distribution by Fω̃ = P(ω̃ ≤ x). Discrete random variables take on a finite number of values ω k , k ∈ K with associated probabilities, I f (ω k ) = P(ω̃ = ω k ) with P k ∈K f (ω k ) = 1 5 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Random Variables ... I Continuous r.v.’s are described by a density function f (ω) I Probability of ω being in an interval [a, b] is Z b P(a ≤ ω̃ ≤ b) = f (ω̃)d ω̃ Z a b = dF (ω̃) a = F (b) − F (a) I Contrary to the discrete case, P(ω̃ = x) = 0 6 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I More I Expected value of ω̃ is I I I P Discrete case: µ = Eω̃ = kR∈K ω k f (ω k ) R∞ ∞ Continuous case: µ = Eω̃ = −∞ ω̃f (ω̃)d ω̃ = −∞ ω̃dF (ω̃) Variance of ω̃ is Var(ω̃)= E[(ω̃ − µ)2 ] 7 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Sources of Uncertainty I Since the real world is uncertain, it becomes imperative to consider uncertainty in decision-making. I Example Sources of Uncertainty: I I I I I I I Market (product, stocks) related. Financial related. Technology related. Competition related. Weather related (e.g. airline rescheduling). Catastrophic events (accidents, war, 9/11, etc.). And many more ... 8 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Example Applications I Manufacturing (supply chain planning) I Transportation (e.g airline industry) I Telecommunications (e.g. network design) I Electricity power generation (e.g. power adequacy planning) I Health care (e.g. patient/resource scheduling) I Agriculture / forestry (e.g. wildfire emergency response) I Finance (e.g. portfolio optimization) And many more ... 9 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Sequential Decision Models I Decision Trees I Graphical representation of the decision process. I 2 - Decision “event". I ° - Uncertainty “event". I → - Time, progressing from left to right. I Random event is a point at which information is revealed/provided. 10 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision Trees I Decision trees help to keep track of interplay between decisions and random events. I Decisions that follow “info" can adapt to it. I Decisions that precede it cannot. Decision Observe Decision ... I In general, it is “best" to delay the decision as long as possible: most flexible for adapting to “info". I There is no advantage to delay if no “info" is anticipated. 11 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Models I Statistical Decision Theory (SDT)- Wald[1950]... I I Determine best levels of variables that affect the outcome of an experiment. x ∈ X , ω ∈ Ω, associated distribution F (ω), and reward r (x, ω), the basic problem is: Z Max Eω [r (x, ω)|F ] = Max r (x, ω)dF (ω). (1) x∈X I x∈X ω Problem (1) is the fundamental form of stochastic programming. Underlying assumptions lead to major differences between the fields. 12 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Statistical Decision Theory (SDT)- Wald[1950]... I Emphasis is on: I I Changes in F to some updated distribution F̂x that depends on a partial choice of x and some observations of ω. Assumes that this part of analysis dominates any solution procedure, as when X is a small finite set that can easily be enumerated. 13 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Decision Analysis (DA)- Raiffa [1968]... I I Particular part of SDT. Emphasis is on: I I I Acquiring information about possible outcomes. Evaluating the utility associated with possible outcomes. Defining a limited set of possible outcomes usually in the form of a decision tree. 14 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Dynamic Programming (DP) and Markov Decision Processes (MDP) - e.g. Bellman [1957], Ross [1983]... I I I I I Search for optimal actions to take at discrete points in time. Actions influenced by random outcomes and carry one from some state at some stage t to another state at stage t + 1. Emphasizes identifying finite or at least, low-dimensional state and actions spaces in assuming some Markovian structure (actions and outcomes depend on current state). Backward recursion resulting in an optimal decision associated with each state. For infinite horizon problems use discounting; finding stationary policy. 15 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Optimal Stochastic Control I I I I Models often similar to stochastic programming models. Problem dimensions are lower. Emphasizes control rules. More restrictive constraint assumptions. 16 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Stochastic Programming (SP) I I I I I I Generally assumes that difficulties in finding the form of r and changes in F as a function of actions are small compared to finding expectations with known F and optimal x with all other info known. Emphasis is on finding a solution after a suitable problem statement in the form (1) has been found. Basically generalizations of deterministic mathematical programs in which uncontrollable data are not known with certainty. Note: The "certainty" assumption in linear programming (LP) is violated! Stochastic linear programming (SLP) deals with linear programs with random data (course focus). Stochastic mixed-integer programming (SMIP) deals with mixed integer programs with random data. Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Decision-Making Models... I Stochastic Programming ... I Key features: I I I I I Many decision variables with many potential values. Discrete time periods for decisions. Use of expectation (and other risk measures) functionals for objectives. Known or partially known probability distributions. The relative importance of these features contrasts with the other models. 18 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming LP Review: You know this already! 19 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming LP Review 20 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming LP Review 21 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Consider the following linear program (LP): Min c > x s.t. Ax ≥ b, (2a) (2b) Tx ≥ r , x ≥ 0, (2c) (2d) where, I I I I I I x ∈ <n+1 is the decision variable vector c ∈ <n1 is the cost vector A ∈ <m1 ×n1 is the constraint matrix b ∈ <m1 is the RHS vector. T ∈ <m2 ×n1 is the technology matrix r ∈ <m2 is the RHS vector. 22 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I In LP we deal with the following concepts: I I Feasibility Optimality I Both these concepts are clear. In fact, sensitivity analysis in LP deals with these two concepts. I But suppose (T , r ) contains random variables (T̃ , r̃ ). What should we do? 23 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Assumptions: I I Real value of (T , r ) is not known Uncertainty if expressed by probability distribution (e.g. "scenarios"): P(T̃ , r̃ ) : P(T ω , r ω ) = pω , ω ∈ Ω. This can also be expressed as: P(T̃ , r̃ ) : P(T s , r s ) = ps , s = 1, ..., S, I where, S = |Ω| Probability distribution known (available data, experts, forecasting/prediction models, etc.) 24 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming I Decision-making under uncertainty using SP: I I Find x “here-and-now" without knowing the real value of (T , r ), but knowing its probability distribution Tx ≥ r can be interpreted as a goal constraint to be specified more precisely 25 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches I Expected Value Solution I Replace T̃ x ≥ r̃ with E[T̃ ]x ≥ E[r̃ ] = T̄ x ≥ r̄ , where, P P T̄ = s ps T s and r̄ = s ps r s . Then the problem can be given as follows: Min c > x s.t. Ax ≥ b T̄ x ≥ r̄ x ≥ 0. I I I (3) Advantage: Simple model - deterministic LP. Disadvantage: Risk is not taken care of since T s x ≥ r s for some scenarios only! What else can you do here? I Apply sensitivity analysis: Still poor model of decision-making under uncertainty [See Higle and Wallace, 2003]. 26 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I ‘Fat’ Solution I Replace T̃ x ≥ r̃ with T s x ≥ r s , s = 1, ..., S. Then the problem can be given as follows: Min c > x s.t. Ax ≥ b T s x ≥ r s , s = 1, ..., S x ≥ 0. I I (4) Advantage: Deterministic LP. Disadvantage: This is a conservative, expensive, restrictive model: often no feasible solution exists. Yields overly conservative solutions driven by extreme events, no matter how rare! 27 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Scenario Analysis I Solve Min c > x s s.t. Ax ≥ b T sx s ≥ r s x s ≥ 0. I I I (5) for every scenario (T s , r s ), s = 1, ..., S. Get solutions x s , s = 1, ..., S. Find an overall solution “based on the scenario solutions". Advantage: Each scenario solution is a deterministic LP; improvement over the expected value approach; very popular approach. Disadvantage: How do you find an overall solution? 28 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Chance (Probabilistic) Constraints I Replace T̃ x ≥ r̃ with P{T̃ x ≥ r̃ } ≥ α, for some specified reliability level α ∈ (0.5, 1). Then the problem can be given as follows: Min c > x s.t. Ax ≥ b P{T̃ x ≥ r̃ } ≥ α, x ≥ 0. I I (6) Advantage: Risk is taken care of explicitly (1 − α is maximal acceptable risk) Disadvantage: Difficult to compute; discrete distributions may lead to MIP model; in general, possibly nonconvex model. 29 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Two-Stage Recourse Model I Introduce explicitly corrective actions. Replace T̃ x ≥ r̃ with T̃ x + Wy ≥ r̃ , <n+2 I I where y ∈ is the decision vector of a second-stage LP problem. The value of y depends on the realization of (T̃ , r̃ ) Penalize corrective actions, called recourse actions in SP. Minimize total expected costs. Decision x Stage 1 Observe Uncertainty Decision y = ys Stage 2 30 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Two-Stage SLP with Recourse Model (discrete distribution): Min c > x + S X ps q > y s (7a) s=1 s.t. Ax ≥ b, s s T x + Wy ≥ r s , x ≥ 0, y s ≥ 0, s = 1, ..., S, (7b) (7c) (7d) with q unit recourse costs. Objective = c > x + expected recourse costs. 31 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Two-Stage SP with Recourse Model ... I I I Advantage: Risk is taken care of explicitly (expected recourse costs); large-scale LP model. Disadvantage: model may be too large to solve, e.g. 10 independent random variables, 6 realizations each ⇒ S = 610 ≈ 60, 000, 000! SP dimensions: matrix A is (m1 × n1 ), W is (m2 × n2 ), SP has (n1 + n2 S) decision variables (m1 + m2 S) constraints. Large-scale model ⇒ DECOMPOSITION approach to solve! 32 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I A general two-stage SP with recourse model can be written as follows: Min c > x + Eω̃ [f (x, ω̃)] s.t. Ax ≥ b x ≥ 0, (8) where for any realization ω of ω̃ (defined on a probability space (Ω, A, P)) we have Min q(ω)> y s.t. W (ω)y ≥ r (ω) − T (ω)x y ≥ 0. (9) 33 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Some standard SP terminology: I I Function E[f (x, ω̃)]: Expected recourse function Matrix W (ω): Recourse matrix I I I I I I Matrix W (ω) = W fixed: Fixed recourse Matrix W (ω) random: Random recourse Matrix W = [I, −I]: Simple recourse If (f (x, ω̃)) < ∞ w.p.1 ∀x ∈ <n1 : Complete recourse If (f (x, ω̃)) < ∞ w.p.1 ∀x ∈ X , X = {x ∈ <n1 |Ax ≥ b} Relatively complete recourse We will talk more about these later 34 / 35 Probability Spaces Applications Decision-Making Models Linear Programming (LP) Review From LP to Stochastic Programming Approaches... I Next ... I Stochastic programming modeling examples 35 / 35