Table of Contents
 DECIS
 GAMS/DECIS
 Description of GAMS/DECIS Options
 Appendix A  GAMS/DECIS Illustrative Examples
 Appendix B  Error Messages
 DECIS License and Warranty
DECIS
Introduction
DECIS is a system for solving largescale stochastic programs, i.e. programs that include parameters (coefficients and righthand sides) that are not known with certainty, but are assumed to be known by their probability distribution. It employs Benders decomposition and advanced Monte Carlo sampling techniques. DECIS includes a variety of solution strategies, such as solving the universe problem, the expected value problem, Monte Carlo sampling within the Benders decomposition algorithm, and Monte Carlo presampling. When using Monte Carlo sampling the user has the option of employing crude Monte Carlo without variance reduction techniques, or using as variance reduction techniques importance sampling or control variates, based on either an additive or a multiplicative approximation function. Presampling is limited to using crude Monte Carlo only.
For solving linear and nonlinear programs (master and subproblems arising from the decomposition) DECIS interfaces with MINOS or CPLEX. MINOS, see Murtagh and Saunders (1983) [141] , is a stateoftheart solver for largescale linear and nonlinear programs, and CPLEX is one of the fastest linear programming solvers available.
For details about the DECIS system consult the DECIS User's Guide see Infanger (1997) [99] . It includes a comprehensive mathematical description of the methods used by DECIS. In this Guide we concentrate on how to use DECIS directly from GAMS, and especially on how to model stochastic programs using the GAMS/DECIS interface. First, however, in section What DECIS Can Do we give a brief description of what DECIS can do and what solution strategies it uses. This description has been adapted from the DECIS User's Guide. In section GAMS/DECIS we discuss in detail how to set up a stochastic problem using GAMS/DECIS and give a description of the parameter setting and outputs obtained. In Appendix A  GAMS/DECIS Illustrative Examples we show the GAMS/DECIS formulation of two illustrative examples (APL1P and APL1PC) discussed in the DECIS User's Guide. A list of DECIS error messages are represented in Appendix B  Error Messages.
What DECIS Can Do
DECIS solves twostage stochastic linear programs with recourse:
\begin{eqnarray*} \min\ z = c x + E f^\omega y^\omega & & \\ s/t \quad \quad Ax \quad \quad \quad \quad &=&b \\  B^\omega x + \quad D^\omega y^\omega &= &d^\omega \\ x, \quad \quad \quad y^\omega &\geq& 0,\quad \omega\in\Omega. \\ \end{eqnarray*}
where \(x\) denotes the firststage, \(y^\omega\) the secondstage decision variables, \(c\) represents the firststage and \(f^\omega\) the secondstage objective coefficients, \(A\), \(b\) represent the coefficients and right hand sides of the firststage constraints, and \(B^\omega\), \(D^\omega\), \(d^\omega\) represent the parameters of the secondstage constraints, where the transition matrix \(B^\omega\) couples the two stages. In the literature \(D^\omega\) is often referred to as the technology matrix or recourse matrix. The first stage parameters are known with certainty. The second stage parameters are random parameters that assume outcomes labeled \(\omega\) with probability \(p(\omega)\), where \(\Omega\) denotes the set of all possible outcome labels.
At the time the firststage decision \(x\) has to be made, the secondstage parameters are only known by their probability distribution of possible outcomes. Later, after \(x\) has already been determined, an actual outcome of the secondstage parameters will become known, and the secondstage decision \(y^\omega\) is made based on knowledge of the actual outcome \(\omega\). The objective is to find a feasible decision \(x\) that minimizes the total expected costs, the sum of firststage costs and expected secondstage costs.
For discrete distributions of the random parameters, the stochastic linear program can be represented by the corresponding equivalent deterministic linear program:
\[ \begin{array}{rcrcrcrcccrcl} \min\ z &=& cx &+& p^1 fy^1 &+& p^2fy^2 &+&\cdots&+&p^Wfy^W & & \\ s/t && Ax & & & & & & & & &=& b \\ &&B^1x&+&Dy^1 & & & & & & &=& d^1 \\ &&B^2x& & &+&Dy^2 & & & & &=& d^2 \\ &&\vdots& & & & & &\ddots& & & &\vdots \\ &&B^Wx& & & & & & &+&Dy^W &=&d^W \\ &&x, & & y^1, & &y^2, & &\ldots,& & y^W &\geq& 0, \\ \end{array} \]
which contains all possible outcomes \(\omega \in \Omega\). Note that for practical problems \(W\) is very large, e.g., a typical number could be \(10^{20}\), and the resulting equivalent deterministic linear problem is too large to be solved directly.
In order to see the twostage nature of the underlying decision making process the folowing representation is also often used:
\begin{eqnarray*} \min\ cx \quad + & E\ z^\omega(x) & \\ Ax \qquad & & = b \\ x \qquad & &\ge 0 \\ \end{eqnarray*}
where
\begin{eqnarray*} z^\omega(x) = \min\ f^\omega y^\omega & & \\ D^\omega y^\omega &=& d^\omega + B^\omega x \\ y^\omega &\geq& 0,\ \omega\in\Omega = \{1,2,\ldots,W \}. \\ \end{eqnarray*}
DECIS employs different strategies to solve twostage stochastic linear programs. It computes an exact optimal solution to the problem or approximates the true optimal solution very closely and gives a confidence interval within which the true optimal objective lies with, say, 95% confidence.
Representing Uncertainty
It is favorable to represent the uncertain secondstage parameters in a structure. Using \(V=(V_1, \ldots, V_h)\) an \(h\)dimensional independent random vector parameter that assumes outcomes \(v^\omega = (v_1, \ldots, v_h)^\omega\) with probability \(p^\omega = p(v^\omega)\), we represent the uncertain secondstage parameters of the problem as functions of the independent random parameter \(V\):
Each component \(V_i\) has outcomes \(v_i^{\omega_i}\), \(\omega_i \in \Omega_i\), where \(\omega_i\) labels a possible outcome of component \(i\), and \(\Omega_i\) represents the set of all possible outcomes of component \(i\). An outcome of the random vector
consists of \(h\) independent component outcomes. The set
represents the crossing of sets \(\Omega_i\). Assuming each set \(\Omega_i\) contains \(W_i\) possible outcomes, \(\Omega_i = W_i\), the set \(\Omega\) contains \(W = \prod W_i\) elements, where \(\Omega = W\) represents the number of all possible outcomes of the random vector \(V\). Based on independence, the joint probability is the product
Let \(\eta\) denote the vector of all secondstage random parameters, e.g., \(\eta = \hbox{vec}(f, B, D, d)\). The outcomes of \(\eta\) may be represented by the following general linear dependency model:
where \(H\) is a matrix of suitable dimensions. DECIS can solve problems with such general linear dependency models.
Solving the Universe Problem
We refer to the universe problem if we consider all possible outcomes \(\omega \in \Omega\) and solve the corresponding problem exactly. This is not always possible, because there may be too many possible realizations \(\omega \in \Omega\). For solving the problem DECIS employs Benders decomposition, splitting the problem into a master problem, corresponding to the firststage decision, and into subproblems, one for each \(\omega \in \Omega\), corresponding to the secondstage decision. The details of the algorithm and techniques used for solving the universe problem are discussed in The DECIS User's Manual.
Solving the universe problem is referred to as strategy 4. Use this strategy only if the number of universe scenarios is reasonably small. There is a maximum number of universe scenarios DECIS can handle, which depends on your particular resources.
Solving the Expected Value Problem
The expected value problem results from replacing the stochastic parameters by their expectation. It is a linear program that can also easily be solved by employing a solver directly. Solving the expected value problem may be useful by itself (for example as a benchmark to compare the solution obtained from solving the stochastic problem), and it also may yield a good starting solution for solving the stochastic problem. DECIS solves the expected value problem using Benders decomposition. The details of generating the expected value problem and the algorithm used for solving it are discussed in the DECIS User's Manual. To solve the expected value problem choose strategy 1.
Using Monte Carlo Sampling
As noted above, for many practical problems it is impossible to obtain the universe solution, because the number of possible realizations \(\Omega\) is way too large. The power of DECIS lies in its ability to compute excellent approximate solutions by employing Monte Carlo sampling techniques. Instead of computing the expected cost and the coefficients and the righthand sides of the Benders cuts exactly (as it is done when solving the universe problem), DECIS, when using Monte Carlo sampling, estimates the quantities in each iteration using an independent sample drawn from the distribution of the random parameters. In addition to using crude Monte Carlo, DECIS uses importance sampling or control variates as variance reduction techniques.
The details of the algorithm and the different techniques used are described in the DECIS User's Manual. You can choose crude Monte Carlo, referred to as strategy 6, Monte Carlo importance sampling, referred to as strategy 2, or control variates, referred to as strategy 10. Both Monte Carlo importance sampling and control variates have been shown for many problems to give a better approximation compared to employing crude Monte Carlo sampling.
When using Monte Carlo sampling DECIS computes a close approximation to the true solution of the problem, and estimates a close approximation of the true optimal objective value. It also computes a confidence interval within which the true optimal objective of the problem lies, say with 95% confidence. The confidence interval is based on rigorous statistical theory. An outline of how the confidence interval is computed is given in the DECIS User's Manual. The size of the confidence interval depends on the variance of the secondstage cost of the stochastic problem and on the sample size used for the estimation. You can expect the confidence interval to be very small, especially when you employ importance sampling or control variates as a variance reduction technique.
When employing Monte Carlo sampling techniques you have to choose a sample size (set in the parameter file). Clearly, the larger the sample size the better will be the approximate solution DECIS computes, and the smaller will be the confidence interval for the true optimal objective value. The default value for the sample size is 100. Setting the sample size too small may lead to bias in the estimation of the confidence interval, therefore the sample size should be at least 30.
Monte Carlo Presampling
We refer to presampling when we first take a random sample from the distribution of the random parameters and then generate the approximate stochastic problem defined by the sample. The obtained approximate problem is then solved exactly using decomposition. This is in contrast to the way we used Monte Carlo sampling in the previous section, where we used Monte Carlo sampling in each iteration of the decomposition.
The details of the techniques used for presampling are discussed in the DECIS User's Manual. DECIS computes the exact solution of the sampled problem using decomposition. This solution is an approximate solution of the original stochastic problem. Besides this approximate solution, DECIS computes an estimate of the expected cost corresponding to this approximate solution and a confidence interval within which the true optimal objective of the original stochastic problem lies with, say, 95% confidence. The confidence interval is based on statistical theory, while its size depends on the variance of the secondstage cost of the stochastic problem and on the sample size used for generating the approximate problem. In conjunction with presampling no variance reduction techniques are currently implemented.
Using Monte Carlo presampling you have to choose a sample size. Clearly, the larger the sample size you choose, the better will be the solution DECIS computes, and the smaller will be the confidence interval for the true optimal objective value. The default value for the sample size is 100. Again, setting the sample size too small may lead to a bias in the estimation of the confidence interval, therefore the sample size should be at least 30.
To use Monte Carlo presampling choose strategy 8.
Regularized Decomposition
When solving practical problems, the number of Benders iterations can be quite large. In order to control the decomposition, with the hope to reduce the iteration count and the solution time, DECIS makes use of regularization. When employing regularization, an additional quadratic term is added to the objective of the master problem, representing the square of the distance between the best solution found so far (the incumbent solution) and the variable \(x\). Using this term, DECIS controls the distance between solutions in successive decomposition iterations.
To enable regularization you have to set the corresponding parameter. You also have to choose the value of the constant rho in the regularization term. The default is regularization disabled. Details of how DECIS carries out regularization are represented in the DECIS User's Manual.
Regularization is only implemented when using MINOS as the optimizer for solving subproblems. Regularization has proven to be helpful for problems that need a large number of Benders iteration when solved without regularization. Problems that need only a small number of Benders iterations without regularization are not expected to improve much with regularization, and may need even more iterations with regularization than without.
GAMS/DECIS
GAMS stands for General Algebraic Modeling Language, and is one of the most widely used modeling languages. Using DECIS directly from GAMS spares you from worrying about all the details of the input formats. It makes the problem formulation much easier but still gives you almost all the flexibility of using DECIS directly.
The link from GAMS to DECIS has been designed in such a way that almost no extensions to the GAMS modeling language were necessary for carrying out the formulation and solution of stochastic programs. In a future release of GAMS, however, additions to the language are planned that will allow you to model stochastic programs in an even more elegant way.
Setting up a Stochastic Program Using GAMS/DECIS
The interface from GAMS to DECIS supports the formulation and solution of stochastic linear programs. DECIS solves them using twostage decomposition. The GAMS/DECIS interface resembles closely the structure of the SMPS (stochastic mathematical programming interface) discussed in the DECIS User's Manual. The specification of a stochastic problem using GAMS/DECIS uses the following components:
 the deterministic (core) model,
 the specification of the decision stages,
 the specification of the random parameters, and
 setting DECIS as the optimizer to be used.
Starting with the Deterministic Model
The core model is a deterministic linear program where all random parameters are replaced by their mean or by a particular realization. One could also see it as a GAMS model without any randomness. It could be a deterministic model that you have, that you intend to expand to a stochastic one. Using DECIS with GAMS allows you to easily extend a deterministic linear programming model to a stochastic one. For example, the following GAMS model represents a deterministic version of the electric power expansion planning illustrative example discussed in Infanger (1994).
* APL1P test model * Dr. Gerd Infanger, November 1997 * Deterministic Program set g generators / g1, g2/; set dl demand levels /h, m, l/; parameter alpha(g) availability / g1 0.68, g2 0.64 /; parameter ccmin(g) min capacity / g1 1000, g2 1000 /; parameter ccmax(g) max capacity / g1 10000, g2 10000 /; parameter c(g) investment / g1 4.0, g2 2.5 /; table f(g,dl) operating cost h m l g1 4.3 2.0 0.5 g2 8.7 4.0 1.0; parameter d(dl) demand / h 1040, m 1040, l 1040 /; parameter us(dl) cost of unserved demand / h 10, m 10, l 10 /; free variable tcost total cost; positive variable x(g) capacity of generators; positive variable y(g, dl) operating level; positive variable s(dl) unserved demand; equations cost total cost cmin(g) minimum capacity cmax(g) maximum capacity omax(g) maximum operating level demand(dl) satisfy demand; cost .. tcost =e= sum(g, c(g)*x(g)) + sum(g, sum(dl, f(g,dl)*y(g,dl))) + sum(dl,us(dl)*s(dl)); cmin(g) .. x(g) =g= ccmin(g); cmax(g) .. x(g) =l= ccmax(g); omax(g) .. sum(dl, y(g,dl)) =l= alpha(g)*x(g); demand(dl) .. sum(g, y(g,dl)) + s(dl) =g= d(dl); model apl1p /all/; option lp=minos; solve apl1p using lp minimizing tcost; scalar ccost capital cost; scalar ocost operating cost; ccost = sum(g, c(g) * x.l(g)); ocost = tcost.l  ccost; display x.l, tcost.l, ccost, ocost, y.l, s.l;
Setting the Decision Stages
Next in order to extend a deterministic model to a stochastic one you must specify the decision stages. DECIS solves stochastic programs by twostage decomposition. Accordingly, you must specify which variables belong to the first stage and which to the second stage, as well as which constraints are firststage constraints and which are secondstage constraints. First stage constraints involve only firststage variables, while secondstage constraints involve both first and secondstage variables. You must specify the stage of a variable or a constraint by setting the stage suffix ".STAGE" to either one or two depending on if it is a first or second stage variable or constraint. For example, expanding the illustrative model above by
* setting decision stages x.stage(g) = 1; y.stage(g, dl) = 2; s.stage(dl) = 2; cmin.stage(g) = 1; cmax.stage(g) = 1; omax.stage(g) = 2; demand.stage(dl) = 2;
would make \(\hbox{x(g)}\) firststage variables, \(\hbox{y(g, dl)}\) and \(\hbox{s(dl)}\) secondstage variables, \(\hbox{cmin(g)}\) and \(\hbox{cmax(g)}\) firststage constraints, and \(\hbox{omax(g)}\) and \(\hbox{demand(g)}\) secondstage constraints. The objective is treated separately, you don't need to set the stage suffix for the objective variable and objective equation.
Note that the use of the .stage
variable and equation suffix causes the GAMS scaling facility (i.e. the .scale
suffix) to be unavailable. Stochastic models have to be scaled manually.
Specifying the Stochastic Model
DECIS supports any linear dependency model, i.e., the outcomes of an uncertain parameter in the linear program are a linear function of a number of independent random parameter outcomes. DECIS considers only discrete distributions: you must approximate any continuous distributions by discrete ones. The number of possible realizations of the discrete random parameters determines the accuracy of the approximation. A special case of a linear dependency model arises when you have only independent random parameters in your model. In this case the independent random parameters are mapped one to one into the random parameters of the stochastic program. We will present the independent case first and then expand to the case with linear dependency. According to setting up a linear dependency model we present the formulation in GAMS by first defining independent random parameters and then defining the distributions of the uncertain parameters in your model.
Specifying Independent Random Parameters
There are of course many different ways you can set up independent random parameters in GAMS. In the following we show one possible way that is generic and thus can be adapted for different models. The setup uses the set \(\hbox{stoch}\) for labeling outcome named "out" and probability named "pro" of each independent random parameter. In the following we show how to define an independent random parameter, say, \(\hbox{v1}\). The formulation uses the set \(\hbox{omega1}\) as driving set, where the set contains one element for each possible realization the random parameter can assume. For example, the set \(\hbox{omega1}\) has four elements according to a discrete distribution of four possible outcomes. The distribution of the random parameter is defined as the parameter \(\hbox{v1}\), a twodimensional array of outcomes "out" and corresponding probability "pro" for each of the possible realizations of the set \(\hbox{omega1}\), "o11", "o12", "o13", and "o14". For example, the random parameter \(\hbox{v1}\) has outcomes of \(1.0, 0.9, 0.5, 0.1\) with probabilities \(0.2, 0.3, 0.4, 0.1\), respectively. Instead of using assignment statements for inputting the different realizations and corresponding probabilities you could also use the table statement. Always make sure that the sum of the probabilities of each independent random parameter adds to one.
* defining independent stochastic parameters set stoch /out, pro /; set omega1 / o11, o12, o13, o14 /; table v1(stoch, omega1) o11 o12 o13 o14 out 1.0 0.9 0.5 0.1 pro 0.2 0.3 0.4 0.1 ;
Random parameter \(\hbox{v1}\) is the first out of five independent random parameters of the illustrative model APL1P, where the first two represent the independent availabilities of the generators \(\hbox{g1}\) and \(\hbox{g2}\) and the latter three represent the independent demands of the demand levels \(\hbox{h}\), \(\hbox{m}\), and \(\hbox{l}\). We also represent the definitions of the remaining four independent random parameters. Note that random parameters \(\hbox{v3}\), \(\hbox{v4}\), and \(\hbox{v5}\) are identically distributed.
set omega2 / o21, o22, o23, o24, o25 /; table v2(stoch, omega2) o21 o22 o23 o24 o25 out 1.0 0.9 0.7 0.1 0.0 pro 0.1 0.2 0.5 0.1 0.1 ; set omega3 / o31, o32, o33, o34 /; table v3(stoch, omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ; set omega4 / o41, o42, o43, o44 /; table v4(stoch,omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ; set omega5 / o51, o52, o53, o54 /; table v5(stoch,omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ;
Defining the Distributions of the Uncertain Parameters in the Model
Having defined the independent stochastic parameters (you may copy the setup above and adapt it for your model), we next define the stochastic parameters in the GAMS model. The stochastic parameters of the model are defined by writing a file, the GAMS stochastic file, using the put facility of GAMS. The GAMS stochastic file resembles closely the stochastic file of the SMPS input format. The main difference is that we use the row, column, bounds, and right hand side names of the GAMS model and that we can write it in free format.
Independent Stochastic Parameters
First we describe the case where all stochastic parameters in the model are independent, see below the representation of the stochastic parameters for the illustrative example APL1P, which has five independent stochastic parameters.
First define the GAMS stochastic file "MODEL.STG" (only the exact name in uppercase letters is supported) and set up GAMS to write to it. This is done by the first two statements. You may want to consult the GAMS manual for how to use put for writing files. The next statement "INDEP DISCRETE" indicates that a section of independent stochastic parameters follows. We then write all possible outcomes and corresponding probabilities for each stochastic parameter by using loop statements. Of course one could also write each line separately, but the loops work nicely. Writing a "*" between the definitions of the independent stochastic parameters is merely for optical reasons and can be omitted.
* defining distributions (writing file MODEL.STG) file stg /MODEL.STG/; put stg; put "INDEP DISCRETE" /; loop(omega1, put "x g1 omax g1 ", v1("out", omega1), " period2 ", v1("pro", omega1) /; ); put "*" /; loop(omega2, put "x g2 omax g2 ", v2("out", omega2), " period2 ", v2("pro", omega2) /; ); put "*" /; loop(omega3, put "RHS demand h ", v3("out", omega3), " period2 ", v3("pro", omega3) /; ); put "*" /; loop(omega4, put "RHS demand m ", v4("out", omega4), " period2 ", v4("pro", omega4) /; ) put "*" /; loop(omega5, put "RHS demand l ", v5("out", omega5), " period2 ", v5("pro", omega5) /; ); putclose stg;
In the example APL1P the first stochastic parameter is the availability of generator \(\hbox{g1}\). In the model the parameter appears as the coefficient of variable \(\hbox{x(g1)}\) in equation \(\hbox{omax(g1)}\). The definition using the put statement first gives the stochastic parameter as the intersection of variable \(\hbox{x(g1)}\) with equation \(\hbox{omax(g1)}\), but without having to type the braces, thus x g1 omax g1, then the outcome v1("out", omega1) and the probability v1("pro", omega1) separated by "period2". The different elements of the statement must be separated by blanks. Since the outcomes and probabilities of the first stochastic parameters are driven by the set \(\hbox{omega1}\) we loop over all elements of the set \(\hbox{omega1}\). We continue and define all possible outcomes for each of the five independent stochastic parameters.
In the example of independent stochastic parameters, the specification of the distribution of the stochasic parameters using the put facility creates the following file "MODEL.STG", which then is processed by the GAMS/DECIS interface:
INDEP DISCRETE x g1 omax g1 1.00 period2 0.20 x g1 omax g1 0.90 period2 0.30 x g1 omax g1 0.50 period2 0.40 x g1 omax g1 0.10 period2 0.10 * x g2 omax g2 1.00 period2 0.10 x g2 omax g2 0.90 period2 0.20 x g2 omax g2 0.70 period2 0.50 x g2 omax g2 0.10 period2 0.10 x g2 omax g2 0.00 period2 0.10 * RHS demand h 900.00 period2 0.15 RHS demand h 1000.00 period2 0.45 RHS demand h 1100.00 period2 0.25 RHS demand h 1200.00 period2 0.15 * RHS demand m 900.00 period2 0.15 RHS demand m 1000.00 period2 0.45 RHS demand m 1100.00 period2 0.25 RHS demand m 1200.00 period2 0.15 * RHS demand l 900.00 period2 0.15 RHS demand l 1000.00 period2 0.45 RHS demand l 1100.00 period2 0.25 RHS demand l 1200.00 period2 0.15
For defining stochastic parameters in the righthand side of the model use the keyword RHS as the column name, and the equation name of the equation which righthand side is uncertain, see for example the specification of the uncertain demands RHS demand h, RHS demand m, and RHS demand l. For defining uncertain bound parameters you would use the keywords UP, LO, or FX, the string bnd, and the variable name of the variable whose upper, lower, or fixed bound is uncertain.
Note all the keywords for the definitions are in capital letters, i.e., "INDEP DISCRETE", "RHS", and not represented in the example "UP", "LO", and "FX".
Note that in GAMS equations, variables may appear in the righthand side, e.g. "EQ.. X+1 =L= 2*Y"
. When the coefficient 2 is a random variable, we need to be aware that GAMS will generate the following LP row X  2*Y =L= 1
. Suppose the probability distribution of this random variable is given by:
set s scenario /pessimistic, average, optimistic/; parameter outcome(s) / pessimistic 1.5 average 2.0 optimistic 2.3 /; parameter prob(s) / pessimistic 0.2 average 0.6 optimistic 0.2 /;
then the correct way of generating the entries in the stochastic file would be:
loop(s, put "Y EQ ",(outcome(s))," PERIOD2 ",prob(s)/; );
Note the negation of the outcome parameter. Also note that expressions in a PUT statement have to be surrounded by parentheses. GAMS reports in the row listing section of the listing file how equations are generated. You are encouraged to inspect the row listing to discover or confirm how coefficients appear in a generated LP row.
Dependent Stochastic Parameters
Next we describe the case of general linear dependency of the stochastic parameters in the model, see below the representation of the stochastic parameters for the illustrative example APL1PCA, which has three dependent stochastic demands driven by two independent stochastic random parameters. First we give the definition of the two independent stochastic parameters, which in the example happen to have two outcomes each.
* defining independent stochastic parameters set stoch /out, pro/; set omega1 / o11, o12 /; table v1(stoch,omega1) o11 o12 out 2.1 1.0 pro 0.5 0.5 ; set omega2 / o21, o22 /; table v2(stoch, omega2) o21 o22 out 2.0 1.0 pro 0.2 0.8 ;
We next define the parameters of the transition matrix from the independent stochastic parameters to the dependent stochastic parameters of the model. We do this by defining two parameter vectors, where the vector \(hm1\) gives the coefficients of the independent random parameter \(v1\) in each of the three demand levels and the vector \(hm2\) gives the coefficients of the independent random parameter \(v2\) in each of the three demand levels.
parameter hm1(dl) / h 300., m 400., l 200. /; parameter hm2(dl) / h 100., m 150., l 300. /;
Again first define the GAMS stochastic file "MODEL.STG" and set GAMS to write to it. The statement BLOCKS DISCRETE indicates that a section of linear dependent stochastic parameters follows.
* defining distributions (writing file MODEL.STG) file stg / MODEL.STG /; put stg; put "BLOCKS DISCRETE" /; scalar h1; loop(omega1, put "BL v1 period2 ", v1("pro", omega1)/; loop(dl, h1 = hm1(dl) * v1("out", omega1); put "RHS demand ", dl.tl:1, " ", h1/; ); ); loop(omega2, put " BL v2 period2 ", v2("pro", omega2) /; loop(dl, h1 = hm2(dl) * v2("out", omega2); put "RHS demand ", dl.tl:1, " ", h1/; ); ); putclose stg;
Dependent stochastic parameters are defined as functions of independent random parameters. The keyword BL labels a possible realization of an independent random parameter. The name besides the BL keyword is used to distinguish between different outcomes of the same independent random parameter or a different one. While you could use any unique names for the independent random parameters, it appears natural to use the names you have already defined above, e.g., \(v1\) and \(v2\). For each realization of each independent random parameter define the outcome of every dependent random parameter (as a function of the independent one). If a dependent random parameter in the GAMS model depends on two or more different independent random parameters the contributions of each of the independent parameters are added. We are therefore in the position to model any linear dependency model. (Note that the class of models that can be accommodated here is more general than linear. The functions, with which an independent random variable contributes to the dependent random variables can be any ones in one argument. As a general rule, any stochastic model that can be estimated by linear regression is supported by GAMS/DECIS.)
Define each independent random parameter outcome and the probability associated with it. For example, the statement starting with BL v1 period2 indicates that an outcome of (independent random parameter) v1 is being defined. The name period2 indicates that it is a secondstage random parameter, and v1("pro", omega1) gives the probability associated with this outcome. Next list all random parameters dependent on the independent random parameter outcome just defined. Define the dependent stochastic parameter coefficients by the GAMS variable name and equation name, or "RHS" and variable name, together with the value of the parameter associated with this realization. In the example, we have three dependent demands. Using the scalar \(h1\) for intermediately storing the results of the calculation, looping over the different demand levels \(dl\) we calculate h1 = hm1(dl) * v1("out", omega1) and define the dependent random parameters as the righthand sides of equation \(demand(dl)\).
When defining an independent random parameter outcome, if the block name is the same as the previous one (e.g., when BL v1 appears the second time), a different outcome of the same independent random parameter is being defined, while a different block name (e.g., when BL v2 appears the first time) indicates that the first outcome of a different independent random parameter is being defined. You must ensure that the probabilities of the different outcomes of each of the independent random parameters add up to one. The loop over all elements of \(omega1\) defines all realizations of the independent random parameter \(v1\) and the loop over all elements of \(omega2\) defines all realizations of the independent random parameter \(v2\).
Note that for the first realization of an independent random parameter, you must define all dependent parameters and their realizations. The values entered serve as a base case. For any other realization of an independent random parameter you only need to define the dependent parameters that have different coefficients than have been defined in the base case. For those not defined in a particular realization, their values of the base case are automatically added.
In the example of dependent stochastic parameters above, the specification of the distribution of the stochastic parameters using the put facility creates the following file "MODEL.STG", which then is processed by the GAMS/DECIS interface:
BLOCKS DISCRETE BL v1 period2 0.50 RHS demand h 630.00 RHS demand m 840.00 RHS demand l 420.00 BL v1 period2 0.50 RHS demand h 300.00 RHS demand m 400.00 RHS demand l 200.00 BL v2 period2 0.20 RHS demand h 200.00 RHS demand m 300.00 RHS demand l 600.00 BL v2 period2 0.80 RHS demand h 100.00 RHS demand m 150.00 RHS demand l 300.00
Again all the keywords for the definitions are in capital letters, i.e., "BLOCKS DISCRETE", "BL", "RHS", and not represented in the example "UP", "LO", and "FX".
Note that you can only define random parameter coefficients that are nonzero in your GAMS model. When setting up the deterministic core model put a nonzero entry as a placeholder for any coefficient that you wish to specify as a stochastic parameter. Specifying a random parameter at the location of a zero coefficient in the GAMS model causes DECIS to terminate with an error message.
Setting DECIS as the Optimizer
After having finished the stochastic definitions you must set DECIS as the optimizer. This is done by issuing the following statements:
* setting DECIS as optimizer * DECISM uses MINOS, DECISC uses CPLEX option lp=decism; apl1p.optfile = 1;
The statement option lp = decism sets DECIS with the MINOS LP engine as the optimizer to be used for solving the stochastic problem. Note that if you do not use DECIS, but instead use any other linear programming optimizer, your GAMS model will still run and optimize the deterministic core model that you have specified. The statement apl1p.optfile = 1 directs GAMS to process the file DECIS.OPT
, in which you may define any DECIS parameters.
Setting Parameter Options in the GAMS Model
The options iteration limit and resource limit can be set directly in your GAMS model file. For example, the following statements
option iterlim = 1000; option reslim = 6000;
limit the number of decomposition iterations to be less than or equal to 1000, and the elapsed time for running DECIS to be less than or equal to 6000 seconds or 100 minutes.
Setting Parameters in the DECIS Options File
In the DECIS options file you can specify parameters regarding the solution algorithm used and control the output of the DECIS program. There is a record for each parameter you want to specify. Each record consists of the value of the parameter you want to specify and the keyword identifying the parameter, separated by a blank character or a comma. You may specify parameters with the following keywords: istrat, nsamples, nzrows, iwrite, ibug, iscratch, ireg, rho, tolben, and tolw in any order. Each keyword can be specified in lower case or upper case text in the format (A10). Since DECIS reads the records in free format you don't have to worry about the format, but some computers require that the text is inputted in quotes. Parameters that are not specified in the parameter file automatically assume their default values. See details of these parameters from Section Description of GAMS/DECIS Options.
Example
In the following example the parameters istrat = 7, nsamples = 200, and nzrows = 200 are specified. All other parameters are set at their default values. DECIS first solves the expected value problem and then the stochastic problem using crude Monte Carlo sampling with a sample size of nsamples = 200. DECIS reserves space for a maximum of nzrows = 50 cuts.
7 "ISTRAT" 200 "NSAMPLES" 50 "NZROWS"
Setting MINOS Parameters in the MINOS Specification File
When you use MINOS as the optimizer for solving the master and the subproblems, you must specify optimization parameters in the MINOS specification file "MINOS.SPC". Each record of the file corresponds to the specification of one parameter and consists of a keyword and the value of the parameter in free format. Records having a "*" as their first character are considered as comment lines and are not further processed. For a detailed description of these parameters, see the MINOS Users' Guide (Murtagh and Saunders (1983) [141] . The following parameters should be specified with some consideration:

AIJ TOLERANCE — Specifies the nonzero tolerance for constraint matrix elements of the problem. Matrix elements \(a_{ij}\) that have a value for which \(a_{ij}\) is less than "AIJ TOLERANCE" are considered by MINOS as zero and are automatically eliminated from the problem. It is wise to specify "AIJ TOLERANCE 0.0 "

SCALE — Specifies MINOS to scale the problem ("SCALE YES") or not ("SCALE NO"). It is wise to specify "SCALE NO".

ROWS — Specifies the number of rows in order for MINOS to reserve the appropriate space in its data structures when reading the problem. "ROWS" should be specified as the number of constraints in the core problem or greater.

COLUMNS — Specifies the number of columns in order for MINOS to reserve the appropriate space in its data structures when reading the problem. "COLUMNS" should be specified as the number of variables in the core problem or greater.

ELEMENTS — Specifies the number of nonzero matrix coefficients in order for MINOS to reserve the appropriate space in its data structures when reading the problem. "ELEMENTS" should be specified as the number of nonzero matrix coefficients in the core problem or greater.
Example
The following example represents typical specifications for running DECIS with MINOS as the optimizer.
BEGIN SPECS PRINT LEVEL 1 LOG FREQUENCY 10 SUMMARY FREQUENCY 10 MPS FILE 12 ROWS 20000 COLUMNS 50000 ELEMENTS 100000 ITERATIONS LIMIT 30000 * FACTORIZATION FREQUENCY 100 AIJ TOLERANCE 0.0 * SCALE NO END OF SPECS
Setting CPLEX Parameters Using System Environment Variables
When you use CPLEX as the optimizer for solving the master and the subproblems, optimization parameters must be specified through system environment variables. You can specify the parameters "CPLEXLICDIR", "SCALELP", "NOPRESOLVE", "ITERLOG", "OPTIMALITYTOL", "FEASIBILIITYTOL", and "DUALSIMPLEX".

CPLEXLICDIR — Contains the path to the CPLEX license directory. For example, on a Unix system with the CPLEX license directory in /usr/users/cplex/cplexlicdir you issue the command setenv CPLEXLICDIR /usr/users/cplex/cplexlicdir.

SCALELP — Specifies CPLEX to scale the master and subproblems before solving them. If the environment variable is not set no scaling is used. Setting the environment variable, e.g., by issuing the command setenv SCALELP yes, scaling is switched on.

NOPRESOLVE — Allows to switch off CPLEX's presolver. If the environment variable is not set, presolve will be used. Setting the environment variable, e.g., by setting setenv NOPRESOLVE yes, no presolve will be used.

ITERLOG — Specifies the iteration log of the CPLEX iterations to be printed to the file "MODEL.CPX". If you do not set the environment variable no iteration log will be printed. Setting the environment variable, e.g., by setting setenv ITERLOG yes, the CPLEX iteration log is printed.

OPTIMALITYTOL — Specifies the optimality tolerance for the CPLEX optimizer. If you do not set the environment variable the CPLEX default values are used. For example, setting setenv OPTIMALITYTOL 1.0E7 sets the CPLEX optimality tolerance to 0.0000001.

FEASIBILIITYTOL — Specifies the feasibility tolerance for the CPLEX optimizer. If you do not set the environment variable the CPLEX default values are used. For example, setting setenv FEASIBILITYTOL 1.0E7 sets the CPLEX optimality tolerance to 0.0000001.

DUALSIMPLEX — Specifies the dual simplex algorithm of CPLEX to be used. If the environment variable is not set the primal simplex algorithm will be used. This is the default and works beautifully for most problems. If the environment variable is set, e.g., by setting setenv DUALSIMPLEX yes, CPLEX uses the dual simplex algorithm for solving both master and subproblems.
GAMS/DECIS Output
After successfully having solved a problem, DECIS returns the objective, the optimal primal and optimal dual solution, the status of variables (if basic or not), and the status of equations (if binding or not) to GAMS. In the case of firststage variables and equations you have all information available in GAMS, as you would with any other solver, but instead of obtaining the optimal values for a deterministic core problem you actually have the optimal values for the stochastic problem. However, for secondstage variables and constraints the e xpected values of the optimal primal and optimal dual solution are reported. This saves space and is useful for the calculation of risk measures. However, the information as to what the optimal primal and dual solutions were in the different scenarios of the stochastic programs is not reported back to GAMS. In a future release of the GAMS/DECIS interface the GAMS language is planned to be extended to handle the scenario secondstage optimal primal and dual values at least for selected variables and equations.
While running, DECIS outputs important information about the progress of the execution to your computer screen. After successfully solving a problem, DECIS also outputs its optimal solution to the solution output file "MODEL.SOL". The debug output file "MODEL.SCR" contains important information about the optimization run, and the optimizer output files "MODEL.MO" (when using DECIS with MINOS) or "MODEL.CPX" (when using DECIS with CPLEX) contain solution output from the optimizer used. In the DECIS User's Guide you find a detailed discussion of how to interpret the screen output, the solution report and the information in the output files.
The Screen Output
The output to the screen allows you to observe the progress of a DECIS run. After the program logo and the copyright statement, you see four columns of output being written to the screen as long as the program proceeds. The first column (from left to right) represents the iteration count, the second column the lower bound (the optimal objective of the master problem), the third column the best upper bound (exact value or estimate of the total expected cost of the best solution found so far), and the fourth column the current upper bound (exact value or estimate of the total expected cost of current solution). After successful completion, DECIS quits with "Normal Exit", otherwise, if an error has been encountered, the programs stops with the message "Error Exit".
Example
When solving the illustrative example APL1P using strategy 5, we obtain the following report on the screen:
T H E D E C I S S Y S T E M Copyright (c) 1989  1999 by Dr. Gerd Infanger All rights reserved. iter lower best upper current upper 0 0.9935E+06 1 0.4626E+06 0.2590E+05 0.2590E+05 2 0.2111E+05 0.2590E+05 0.5487E+06 3 0.2170E+05 0.2590E+05 0.2697E+05 4 0.2368E+05 0.2384E+05 0.2384E+05 5 0.2370E+05 0.2384E+05 0.2401E+05 6 0.2370E+05 0.2370E+05 0.2370E+05 iter lower best upper current upper 6 0.2370E+05 7 0.2403E+05 0.2470E+05 0.2470E+05 8 0.2433E+05 0.2470E+05 0.2694E+05 9 0.2441E+05 0.2470E+05 0.2602E+05 10 0.2453E+05 0.2470E+05 0.2499E+05 11 0.2455E+05 0.2470E+05 0.2483E+05 12 0.2461E+05 0.2467E+05 0.2467E+05 13 0.2461E+05 0.2467E+05 0.2469E+05 14 0.2461E+05 0.2465E+05 0.2465E+05 15 0.2463E+05 0.2465E+05 0.2467E+05 16 0.2463E+05 0.2465E+05 0.2465E+05 17 0.2464E+05 0.2465E+05 0.2465E+05 18 0.2464E+05 0.2464E+05 0.2464E+05 19 0.2464E+05 0.2464E+05 0.2464E+05 20 0.2464E+05 0.2464E+05 0.2464E+05 21 0.2464E+05 0.2464E+05 0.2464E+05 22 0.2464E+05 0.2464E+05 0.2464E+05 Normal Exit
The Solution Output File
The solution output file contains the solution report from the DECIS run. Its name is "MODEL.SOL". The file contains the best objective function value found, the corresponding values of the firststage variables, the corresponding optimal secondstage cost, and a lower and an upper bound on the optimal objective of the problem. In addition, the number of universe scenarios and the settings for the stopping tolerance are reported. In the case of using a deterministic strategy for solving the problem, exact values are reported. When using Monte Carlo sampling, estimated values, their variances, and the sample size used for the estimation are reported. Instead of exact upper and lower bounds, probabilistic upper and lower bounds, and a 95% confidence interval, within which the true optimal solution lies with 95% confidence, are reported. A detailed description of the solution output file can be found in the DECIS User's Guide.
The Debug Output File
The debug output file contains the standard output of a run of DECIS containing important information about the problem, its parameters, and its solution. It also contains any error messages that may occur during a run of DECIS. In the case that DECIS does not complete a run successfully, the cause of the trouble can usually be located using the information in the debug output file. If the standard output does not give enough information you can set the debug parameter ibug in the parameter input file to a higher value and obtain additional debug output. A detailed description of the debug output file can be found in the DECIS User's Guide.
The Optimizer Output Files
The optimizer output file "MODEL.MO" contains all the output from MINOS when called as a subroutine by DECIS. You can specify what degree of detail should be outputted by setting the appropriate "PRINT LEVEL" in the MINOS specification file. The optimizer output file "MODEL.CPX" reports messages and the iteration log (if switchwd on using the environment variable) from CPLEX when solving master and sub problems.
Description of GAMS/DECIS Options
DECIS Solver Options
Option  Description  Default 

IBug  Debug output  0 
IReg  Indicator for regularized decomposition  MINOS only  0 
IScratch  Internal unit number for output and debug  17 
IStrat  Defines the solution strategy used  3 
IWrite  Subproblem output  0 
NSamples  Sample size used for the estimation  100 
NZRows  Number of rows reserved for cuts in the master problem  100 
Rho  rho parameter of regularization term in the objective function  1000 
TolBen  Tolerance for stopping the decomposition algorithm  1e7 
TolW  tolerance when writing debug solution output  1e9 
IBug (integer): Debug output ↵
Default:
0
value meaning 0
DECIS does not write any debug output 1
Solution of the master problem on each iteration 2
Value 1 plus scenario index and the optimal objective value for each subproblem solved 3
Value 2 plus information regarding importance sampling 4
Value 3 plus optimal dual variables of the cuts 5
Value 4 plus coeffcients and the righthand side of the cuts 6
Value 5 plus dump of the master problem and the subproblem in MPS format
IReg (boolean): Indicator for regularized decomposition  MINOS only ↵
Default:
0
IScratch (integer): Internal unit number for output and debug ↵
Range: {
1
, ..., ∞}Default:
17
IStrat (integer): Defines the solution strategy used ↵
Default:
3
value meaning 1
Solves the expected value problem 2
Solves the stochastic problem using Monte Carlo importance sampling 3
Refers to istrat = 1 plus istrat = 2 4
Solves the stochastic universe problem 5
Refers to istrat = 1 plus istrat = 4 6
Solves the stochastic problem using crude Monte Carlo sampling 7
Refers to istrat = 1 plus istrat = 6 8
Solves the stochastic problem using Monte Carlo presampling 9
Refers to istrat = 1 plus istrat = 8 10
Solves the stochastic problem using control variates 11
Refers to istrat = 1 plus istrat = 10
IWrite (integer): Subproblem output ↵
Default:
0
value meaning 0
No optimizer output is written 1
Optimizer output is written to the file
NSamples (integer): Sample size used for the estimation ↵
Range: {
30
, ..., ∞}Default:
100
NZRows (integer): Number of rows reserved for cuts in the master problem ↵
Range: {
1
, ..., ∞}Default:
100
Rho (real): rho parameter of regularization term in the objective function ↵
Default:
1000
TolBen (real): Tolerance for stopping the decomposition algorithm ↵
Default:
1e7
TolW (real): tolerance when writing debug solution output ↵
Default:
1e9
Appendix A  GAMS/DECIS Illustrative Examples
Example APL1P
* APL1P test model * Dr. Gerd Infanger, November 1997 set g generators /g1, g2/; set dl demand levels /h, m, l/; parameter alpha(g) availability / g1 0.68, g2 0.64 /; parameter ccmin(g) min capacity / g1 1000, g2 1000 /; parameter ccmax(g) max capacity / g1 10000, g2 10000 /; parameter c(g) investment / g1 4.0, g2 2.5 /; table f(g,dl) operating cost h m l g1 4.3 2.0 0.5 g2 8.7 4.0 1.0; parameter d(dl) demand / h 1040, m 1040, l 1040 /; parameter us(dl) cost of unserved demand / h 10, m 10, l 10 /; free variable tcost total cost; positive variable x(g) capacity of generators; positive variable y(g, dl) operating level; positive variable s(dl) unserved demand; equations cost total cost cmin(g) minimum capacity cmax(g) maximum capacity omax(g) maximum operating level demand(dl) satisfy demand; cost .. tcost =e= sum(g, c(g)*x(g)) + sum(g, sum(dl, f(g,dl)*y(g,dl))) + sum(dl,us(dl)*s(dl)); cmin(g) .. x(g) =g= ccmin(g); cmax(g) .. x(g) =l= ccmax(g); omax(g) .. sum(dl, y(g,dl)) =l= alpha(g)*x(g); demand(dl) .. sum(g, y(g,dl)) + s(dl) =g= d(dl); model apl1p /all/; * setting decision stages x.stage(g) = 1; y.stage(g, dl) = 2; s.stage(dl) = 2; cmin.stage(g) = 1; cmax.stage(g) = 1; omax.stage(g) = 2; demand.stage(dl) = 2; * defining independent stochastic parameters set stoch /out, pro /; set omega1 / o11, o12, o13, o14 /; table v1(stoch, omega1) o11 o12 o13 o14 out 1.0 0.9 0.5 0.1 pro 0.2 0.3 0.4 0.1 ; set omega2 / o21, o22, o23, o24, o25 /; table v2(stoch, omega2) o21 o22 o23 o24 o25 out 1.0 0.9 0.7 0.1 0.0 pro 0.1 0.2 0.5 0.1 0.1 ; set omega3 / o31, o32, o33, o34 /; table v3(stoch, omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ; set omega4 / o41, o42, o43, o44 /; table v4(stoch,omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ; set omega5 / o51, o52, o53, o54 /; table v5(stoch,omega1) o11 o12 o13 o14 out 900 1000 1100 1200 pro 0.15 0.45 0.25 0.15 ; * defining distributions file stg /MODEL.STG/; put stg; put "INDEP DISCRETE" /; loop(omega1, put "x g1 omax g1 ", v1("out", omega1), " period2 ", v1("pro", omega1) /; ); put "*" /; loop(omega2, put "x g2 omax g2 ", v2("out", omega2), " period2 ", v2("pro", omega2) /; ); put "*" /; loop(omega3, put "RHS demand h ", v3("out", omega3), " period2 ", v3("pro", omega3) /; ); put "*" /; loop(omega4, put "RHS demand m ", v4("out", omega4), " period2 ", v4("pro", omega4) /; ); put "*" /; loop(omega5, put "RHS demand l ", v5("out", omega5), " period2 ", v5("pro", omega5) /; ); putclose stg; * setting DECIS as optimizer * DECISM uses MINOS, DECISC uses CPLEX option lp=decism; apl1p.optfile = 1; solve apl1p using lp minimizing tcost; scalar ccost capital cost; scalar ocost operating cost; ccost = sum(g, c(g) * x.l(g)); ocost = tcost.l  ccost; display x.l, tcost.l, ccost, ocost, y.l, s.l;
Example APL1PCA
* APL1PCA test model * Dr. Gerd Infanger, November 1997 set g generators /g1, g2/; set dl demand levels /h, m, l/; parameter alpha(g) availability / g1 0.68, g2 0.64 /; parameter ccmin(g) min capacity / g1 1000, g2 1000 /; parameter ccmax(g) max capacity / g1 10000, g2 10000 /; parameter c(g) investment / g1 4.0, g2 2.5 /; table f(g,dl) operating cost h m l g1 4.3 2.0 0.5 g2 8.7 4.0 1.0; parameter d(dl) demand / h 1040, m 1040, l 1040 /; parameter us(dl) cost of unserved demand / h 10, m 10, l 10 /; free variable tcost total cost; positive variable x(g) capacity of generators; positive variable y(g, dl) operating level; positive variable s(dl) unserved demand; equations cost total cost cmin(g) minimum capacity cmax(g) maximum capacity omax(g) maximum operating level demand(dl) satisfy demand; cost .. tcost =e= sum(g, c(g)*x(g)) + sum(g, sum(dl, f(g,dl)*y(g,dl))) + sum(dl,us(dl)*s(dl)); cmin(g) .. x(g) =g= ccmin(g); cmax(g) .. x(g) =l= ccmax(g); omax(g) .. sum(dl, y(g,dl)) =l= alpha(g)*x(g); demand(dl) .. sum(g, y(g,dl)) + s(dl) =g= d(dl); model apl1p /all/; * setting decision stages x.stage(g) = 1; y.stage(g, dl) = 2; s.stage(dl) = 2; cmin.stage(g) = 1; cmax.stage(g) = 1; omax.stage(g) = 2; demand.stage(dl) = 2; * defining independent stochastic parameters set stoch /out, pro/; set omega1 / o11, o12 /; table v1(stoch,omega1) o11 o12 out 2.1 1.0 pro 0.5 0.5 ; set omega2 / o21, o22 /; table v2(stoch, omega2) o21 o22 out 2.0 1.0 pro 0.2 0.8 ; parameter hm1(dl) / h 300., m 400., l 200. /; parameter hm2(dl) / h 100., m 150., l 300. /; * defining distributions (writing file MODEL.STG) file stg / MODEL.STG /; put stg; put "BLOCKS DISCRETE" /; scalar h1; loop(omega1, put "BL v1 period2 ", v1("pro", omega1)/; loop(dl, h1 = hm1(dl) * v1("out", omega1); put "RHS demand ", dl.tl:1, " ", h1/; ); ); loop(omega2, put " BL v2 period2 ", v2("pro", omega2) /; loop(dl, h1 = hm2(dl) * v2("out", omega2); put "RHS demand ", dl.tl:1, " ", h1/; ); ); putclose stg; * setting DECIS as optimizer * DECISM uses MINOS, DECISC uses CPLEX option lp=decism; apl1p.optfile = 1; solve apl1p using lp minimizing tcost; scalar ccost capital cost; scalar ocost operating cost; ccost = sum(g, c(g) * x.l(g)); ocost = tcost.l  ccost; display x.l, tcost.l, ccost, ocost, y.l, s.l;
Appendix B  Error Messages
 ERROR in MODEL.STO: kwd, word1, word2 was not matched in first realization of block
The specification of the stochastic parameters is incorrect. The stochastic parameter has not been specified in the specification of the first outcome of the block. When specifying the first outcome of a block always include all stochastic parameters corresponding to the block.  Option word1 word2 not supported
You specified an input distribution in the stochastic file that is not supported. Check the DECIS manual for supported distributions.  Error in time file
The time file is not correct. Check the file MODEL.TIM. Check the DECIS manual for the form of the time file.  ERROR in MODEL.STO: stochastic RHS for objective, row name2
The specification in the stochastic file is incorrect. You attempted to specify a stochastic righthand side for the objective row (row name2). Check file MODEL.STO.  ERROR in MODEL.STO: stochastic RHS in master, row name2
The specification in the stochastic file is incorrect. You attempted to specify a stochastic righthand side for the master problem (row name2). Check file MODEL.STO.  ERROR in MODEL.STO: col not found, name1
The specification in the stochastic file is incorrect. The entry in the stochastic file, name1, is not found in the core file. Check file MODEL.STO.  ERROR in MODEL.STO: invalid col/row combination, (name1/name2)
The stochastic file (MODEL.STO) contains an incorrect specification.  ERROR in MODEL.STO: no nonzero found (in B or D matrix) for col/row (name1, name2)
There is no nonzero entry for the combination of name1 (col) and name2(row) in the Bmatrix or in the Dmatrix. Check the corresponding entry in the stochastic file (MODEL.STO). You may want to include a nonzero coefficient for (col/row) in the core file (MODEL.COR).  ERROR in MODEL.STO: col not found, name2
The column name you specified in the stochastic file (MODEL.STO) does not exist in the core file (MODEL.COR). Check the file MODEL.STO.  ERROR in MODEL.STO: stochastic bound in master, col name2
You specified a stochastic bound on firststage variable name2. Check file MODEL.STO.  ERROR in MODEL.STO: invalid bound type (kwd) for col name2
The bound type, kwd, you specified is invalid. Check file MODEL.STO.  ERROR in MODEL.STO: row not found, name2
The specification in the stochastic file is incorrect. The row name, name2, does not exist in the core file. Check file MODEL.STO.  ERROR: problem infeasible
The problem solved (master or subproblem) turned out to be infeasible. If a subproblem is infeasible, you did not specify the problem as having the property of "complete recourse". Complete recourse means that whatever firststage decision is passed to a subproblem, the subproblem will have a feasible solution. It is the best way to specify a problem, especially if you use a sampling based solution strategy. If DECIS encounters a feasible subproblem, it adds a feasibility cut and continues the execution. If DECIS encounters an infeasible master problem, the problem you specified is infeasible, and DECIS terminates. Check the problem formulation.  ERROR: problem unbounded
The problem solved (master or subproblem) turned out to be unbounded. Check the problem formulation.  ERROR: error code: inform
The solver returned with an error code from solving the problem (master or subproblem). Consult the users' manual of the solver (MINOS or CPLEX) for the meaning of the error code, inform. Check the problem formulation.  ERROR: while reading SPECS file
The MINOS specification file (MINOS.SPC) containes an error. Check the specification file. Consult the MINOS user's manual.  ERROR: reading mps file, mpsfile
The core file mpsfile (i.e., MODEL.COR) is incorrect. Consult the DECIS manual for instructions regarding the MPS format.  ERROR: row 1 of problem is not a free row
The first row of the problem is not a free row (i.e., is not the objective row). In order to make the first row a free row, set the row type to be 'N'. Consult the DECIS manual for the MPS specification of the problem.  ERROR: name not found = nam1, nam2
There is an error in the core file (MODEL.COR). The problem cannot be decomposed correctly. Check the core file and check the model formulation.  ERROR: matrix not in staircase form
The constraint matrix of the problem as specified in core file (MODEL.COR) is not in staircase form. The firststage rows and columns and the secondstage rows and columns are mixed within each other. Check the DECIS manual as to how to specify the core file. Check the core file and change the order of rows and columns.
DECIS License and Warranty
The software, which accompanies this license (the "Software"
) is the property of Gerd Infanger and is protected by copyright law. While Gerd Infanger continues to own the Software, you will have certain rights to use the Software after your acceptance of this license. Except as may be modified by a license addendum, which accompanies this license, your rights and obligations with respect to the use of this Software are as follows:
 You may
 Use one copy of the Software on a single computer,
 Make one copy of the Software for archival purposes, or copy the software onto the hard disk of your computer and retain the original for archival purposes,
 Use the Software on a network, provided that you have a licensed copy of the Software for each computer that can access the Software over that network, item After a written notice to Gerd Infanger, transfer the Software on a permanent basis to another person or entity, provided that you retain no copies of the Software and the transferee agrees to the terms of this agreement.
 You may not
 Copy the documentation, which accompanies the Software,
 Sublicense, rent or lease any portion of the Software,
 Reverse engineer, decompile, disassemble, modify, translate, make any attempt to discover the source code of the Software, or create derivative works from the Software.
Limited Warranty:
Gerd Infanger warrants that the media on which the Software is distributed will be free from defects for a period of thirty (30) days from the date of delivery of the Software to you. Your sole remedy in the event of a breach of the warranty will be that Gerd Infanger will, at his option, replace any defective media returned to Gerd Infanger within the warranty period or refund the money you paid for the Software. Gerd Infanger does not warrant that the Software will meet your requirements or that operation of the Software will be uninterrupted or that the Software will be errorfree.
THE ABOVE WARRANTY IS EXCLUSIVE AND IN LIEU OF ALL OTHER WARRANTIES, WHETHER EXPRESS OR IMPLIED, INCLUDING THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
Disclaimer of Damages:
REGARDLESS OF WHETHER ANY REMEDY SET FORTH HEREIN FAILS OF ITS ESSENTIAL PURPOSE, IN NO EVENT WILL GERD INFANGER BE LIABLE TO YOU FOR ANY SPECIAL, CONSEQUENTIAL, INDIRECT OR SIMILAR DAMAGES, INCLUDING ANY LOST PROFITS OR LOST DATA ARISING OUT OF THE USE OR INABILITY TO USE THE SOFTWARE EVEN IF GERD INFANGER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
IN NO CASE SHALL GERD INFANGER'S LIABILITY EXCEED THE PURCHASE PRICE FOR THE SOFTWARE. The disclaimers and limitations set forth above will apply regardless of whether you accept the Software.
General:
This Agreement will be governed by the laws of the State of California. This Agreement may only be modified by a license addendum, which accompanies this license or by a written document, which has been signed by both you and Gerd Infanger. Should you have any questions concerning this Agreement, or if you desire to contact Gerd Infanger for any reason, please write:
Gerd Infanger, 1590 Escondido Way, Belmont, CA 94002, USA.
 Note
 Copyright © 1989 – 1999 by Gerd Infanger. All rights reserved. The GAMS/DECIS User's Guide is copyrighted and all rights are reserved. Information in this document is subject to change without notice and does not represent a commitment on the part of Gerd Infanger. The DECIS software described in this document is furnished under a license agreement and may be used only in accordance with the terms of this agreement. The DECIS software can be licensed through Infanger Investment Technology, LLC or through Gams Development Corporation.