Solver and modelling support for stochastic programming H.I. Gassmann, Dalhousie University Happy...

Post on 19-Jan-2016

214 views 1 download

transcript

Solver and modelling support for stochastic programming

H.I. Gassmann, Dalhousie University

Happy Birthday András, November 2009

© 2009 H.I. Gassmann

Agenda• Stochastic programs• An Example• Algebraic modelling languages

– GAMS– MPL– AMPL

• Frontline Systems• SMI (Stochastic modelling interface)• Optimization Services• Other software

© 2009 H.I. Gassmann

Stochastic programs• Two-stage recourse problems• Multistage recourse problems• Chance-constrained problems• Continuous distributions

© 2009 H.I. Gassmann

An Example (Ferguson & Dantzig)

0,,

,,1,

,,1,s.t.

Emin,

jjij

jjji

ijij

ij

ij

jjj

jiijij

zyx

Jjzyxd

Iiax

zpxc

d

d

xij = # aircraft of type i assigned to route j

yj = # empty seats on route j

zj = # lost sales on route j

In the original formulation, d is defined by marginals

© 2009 H.I. Gassmann

Words to ponder…• “Simple recourse is so previous millennium”• “Always build your scenario tree prior to calling

the solver”• “No need to bother with special algorithms ―

Cplex on the deterministic equivalent is best”• “Robust optimization is irrelevant”

© 2009 H.I. Gassmann

GAMSSET equip := (E1, E2, E3, E4); route := (R1, R2, R3, R4, R5); scen := (S1*S720);

DATA avail[equip] := (E1,10, E2,19, E3,25, E4,15); price[route] := (R1,13, R2,13, R3, 7, R4, 7, R5, 1); capac[equip,route] := (...); cost [equip,route] := (...); demand[route,scen] := (...); prob[scen] := (...);

DECISION VARIABLES Assign[equip,route]; Empty[route,scen]; LostSales[route,scen];

MODEL MIN totalcost = SUM(equip,route: cost*Assign) + SUM(route,scen: prob*price*LostSales);

SUBJECT TO availability[equip]: SUM(route: Assign) <= avail;

passengers[route,scen]: SUM(equip: capac*Assign) – Empty + LostSales = demand;

© 2009 H.I. Gassmann

GAMS• Solve deterministic equivalent • Two-stage problems can use DECIS• Discrete scenarios only

© 2009 H.I. Gassmann

MPLINDEX aircraft : = (A,B,C,D);

STOCHASTIC

INDEP route := (NL1, NL2, ND0, ND1, NB0) -> (r1,r2,r3,r4,r45);

OUTCOME out5 := 1..5;

EVENT RouteOut[route,out5] := (NL1, 1..5, NL2, 1..2, ND0, 1..5, ND1, 1..5, NB0, 1..3);

PROBABILITIES p[route, out5 in RouteOut] := [NL1, 1, 0.2, NL1, 2, 0.05, ... ];

© 2009 H.I. Gassmann

MPL (continued)RANDOM DATA Demand[route, out5 in RouteOut] := ...

DATA AircraftAvail[aircraft] := ...; TicketPrice[route] := ...; PotentialPass[aircraft,route] := ...; MonthlyCost [aircraft,route] := ...;

DECISION VARIABLES Assign[aircraft,route] -> x WHERE (PotentialPass > 0);

STAGE2 VARIABLES EmptySeats[route] -> y1; TurnedAway[route] -> y2;

MODEL MIN TotalCost = SUM(aircraft,route: MonthlyCost * Assign) + SUM(route: TicketPrice * TurnedAway);

SUBJECT TO AircraftCap[aircraft]: SUM(route: Assign) <= AircraftAvail; PassengerBal[route]: SUM(aircraft: PotentialPass * Assign) + TurnedAway[route] – EmptySeats[route] = Demand[route];

© 2009 H.I. Gassmann

MPL capabilities• Solve deterministic equivalent (any MPL solver)• Built-in decomposition solver (two-stage - Cplex) • Communication with SMI • Planned extensions: nested decomposition

© 2009 H.I. Gassmann

AMPLset aircraft := {A1, A2, A3, A4};set routes := {r1, r2, r3, r4, r5};

param avail{aircraft} >= 0;param ticketPrice{routes} > 0;param potentialPassenger{aircraft, routes} default 0;param MonthlyCost {a in aircraft, r in routes: potentialPassenger[a,r] > 0};

var demand{route} random;

var Assign{a in aircraft, r in routes: potentialPassenger[a,r] > 0} >= 0;var EmptySeats{route} suffix stage 2;var TurnedAway{route} suffix stage 2;

minimize totalCost: sum{a in aircraft, r in routes: potentialPassenger[a,r] > 0} MonthlyCost[a,r]*Assign[a,r] + sum{r in routes} TicketPrice[r]*TurnedAway[r];

subject to AircraftCap{a in aircraft}: sum{r in routes: potentialPassenger[a,r] > 0} Assign[a,r] <= avail[a];

subject to PassengerBalance{r in routes}: sum{a in aircraft: potentialPassenger[a,r] > 0} potentialPassenger[a,r] * Assign[a,r]

+ TurnedAway[r] - EmptySeats[r] = demand[r];

© 2009 H.I. Gassmann

Frontline systems• Robust optimization• Chance constraints• Recourse models• Expected value, VaR, CVar objectives• One or two stages• Continuous distributions (automatic sampling)• Deterministic equivalent or simulation• Example

© 2009 H.I. Gassmann

SMI (Stochastic Modeling Interface)

• Coin-OR project• API for interaction with stochastic models• Scenario-based• Discrete distributions

© 2009 H.I. Gassmann

Optimization Services• The Optimization Services project aims to provide

A set of standards to facilitate communication between modeling languages, solvers, problem analyzers, simulation engines, and registry and discovery services in a distributed computing environment.

© 2009 H.I. Gassmann

Solvers

AML

Corporate databases

User interfaceData inter-

change

© 2009 H.I. Gassmann

What Is Optimization Services (OS)?A set of XML-based standards for representing, among others,

– optimization instances (OSiL, also OSgL, OSnL and OSsL)– optimization results (OSrL and OSaL)– optimization solver options (OSoL)– communication between clients and solvers (OSpL)

• Open source libraries to work with these standards• A robust API for solver algorithms and modeling systems • Support for linear, integer, nonlinear and stochastic programs• A command line executable OSSolverService• Executables OSAmplClient and GAMSLinks for AMPL and GAMS • Utilities to convert MPS files and AMPL nl files into OSiL• Server software that works with Apache Tomcat and Apache Axis

© 2009 H.I. Gassmann

Why a standard interface?

.

.

....

Modelling systems

Solvers

.

.

....

Modelling systems

Solvers

n*m hook-ups n+m hook-ups

© 2009 H.I. Gassmann

Why XML?• Existing parsers to check syntax• Easy to generate automatically• Tree structure naturally mirrors expression trees for

nonlinear functions• Arbitrary precision and name space• Automatic attribute checking (e.g., nonnegativity)• Querying capabilities via XQuery• Encryption standards being developed• Easy integration into broader IT infrastructure

© 2009 H.I. Gassmann

OSiL Schema – Deterministic data

© 2009 H.I. Gassmann

Representation of uncertainty• Explicit event trees

– Scenario formulation– Only record data items that differ from parent scenario

• Implicit trees (distribution-based formulation)– Assume independence between stages– Distributions within stage discrete or continuous

• Probabilistic constraints– Simple chance constraints– Joint chance constraints

© 2009 H.I. Gassmann

OSInstance: In-memory representation

• XML elements correspond to C++ classes• Child elements mapped as member classes

• set(), get() and calculate() methods

class OSInstance{public: OSInstance(); InstanceHeader *instanceHeader; InstanceData *instanceData;}; // class OSInstance

© 2009 H.I. Gassmann

OSoL – OS option language• Solver options• Initializations of variables• System requirements• Job parameters• In-memory representation: OSOption• API: get(), set(), add() methods

© 2009 H.I. Gassmann

OSrL and OSResult• Result of the optimization

– Solution status– Statistics– Value of primal and dual variables

• Can be displayed in a browser

© 2009 H.I. Gassmann

Solver support• All versions of OS download with COIN-OR solvers

– Clp– Cbc– Ipopt– Bonmin– Couenne– Symphony

• Additiona solver support– Cplex– GLPK– Lindo

© 2009 H.I. Gassmann

Future developments• Stochastic programming• Cone programming• Instance modification• Solution analysis

© 2009 H.I. Gassmann

Other software• SLP-IOR• SPInE/SAMPL/SMPL• Fort-SP• DECIS• MSLiP• …

© 2009 H.I. Gassmann

Conclusions• Slow but steady progress• Stochastic programming is more than scenario

trees and deterministic equivalents• SP-aware modeling systems