Professional Documents
Culture Documents
Kinetic Uncertainty Single Chance Constraint
Kinetic Uncertainty Single Chance Constraint
Optimization under uncertainty is considered necessary for robust process design and operation.
In this work, a new approach is proposed to solve a kind of nonlinear optimization problem
under uncertainty, in which some dependent variables are to be constrained with a predefined
probability. Such problems are called optimization under chance constraints. By employment
of the monotony of these variables to one of the uncertain variables, the output feasible region
will be mapped to a region of the uncertain input variables. Thus, the probability of holding the
output constraints can be simply achieved by integration of the probability density function of
the multivariate uncertain variables. Collocation on finite elements is used for the numerical
integration, through which sensitivities of the chance constraints can be computed as well. The
proposed approach is applied to the optimization of two process engineering problems under
various uncertainties.
ξ2
1.0
2.0
0.2
0.3
[1.0 0.5
0.5 1.0 ]
output variable.
In Figure 2, the right and left circles represent the
is not true. If the constraints are related to the safety whole region of Yi and ΞS, respectively. The points are
consideration of a process operation, a joint chance some realization of the variables based on their distri-
constraint may be preferred. Single chance constraints bution functions. A point in ΞS leads to a point in Yi
may be used when some output constraints are more through the projection yi ) F(ξS). Because of the
critical than the other ones, and in this case, different monotony, a point yi can lead to a unique ξS through
probability levels in (6) can be specified. the reverse projection ξS ) F-1(yi). The shaded area in
Based on the above context, the stochastic optimiza- the right circle is the output feasible region Y′i such
tion problem we consider is formulated with (3) as the that
objective function subject to (5) or (6) as the constraints.
Solving the problem means to search for optimal values
i } g R
P{yi e ySP (7)
of the decision variables u for a minimum expected
value of the objective function and meanwhile satisfying
the probabilistic constraints, under the condition that is satisfied, while the shaded area Ξ′S in the left circle
the values of the uncertain variables ξ will be randomly is the region, with the bound ξLS, of the random vari-
distributed in their effective region Ξ. able corresponding to Y′i. It can be easily seen that the
representation
3. Solution Approach
P{ξS e ξLS} g R (8)
We first consider the relaxation of the defined sto-
chastic optimization problem to an NLP problem, so that
is the same as (7). This implies that the probability of
it can be solved by a standard NLP solver like SQP. It
holding the output constraint can be computed by
is well-known that the solution approaches to a non-
integration in the corresponding region of the uncertain
linear problem can be classified into simultaneous
variable. It should be noted that all uncertain variables
approaches, where both the state and the decision
have to be taken into account when computing P{yi e
variables are included in NLP, and sequential ap-
proaches, where an integration step is used to compute i }. In addition, the value of the decision variables
ySP
the state variables x and thus only the decision vari- also has an impact on the projected region. Then the
ables u will be solved by NLP. Because a multivariate bound ξLS will change based on the realization of the
integration in the domain of the uncertain variables is individual uncertain variables ξs (s ) 1, ..., S - 1) and
required for evaluating the objective function and the the value of u, i.e.,
inequality constraints, a sequential approach would be
preferred in the case of stochastic optimization. ξLS ) F-1(ξ1,...,ξS-1,ySP
i ,u) (9)
In a sequential NLP framework, the values of both
the constraints and their sensitivities to the decision and this leads to the following representation:
variables are required. While the calculation of the
objective function (3) has been proposed from previous
studies with a straightforward multivariate integra- i } )
P{yi e ySP
∫-∞∞‚‚‚∫-∞∞∫-∞ξ F(ξ1,...,ξS-1,ξS) dξS dξS-1 ... dξ1
L
tion,6 no reports on the calculation of the values of the S
(10)
probability for the nonlinear chance constraints (5) and
(6) have been found in the open literature. The challenge
lies in the difficulty in determining the output con- The right-hand side of (10) is a multivariate integration
straints by a nonlinear propagation of the uncertain in the whole region of Ξ except for the part over the
input variables ξ. Therefore, the key problem is the limiting value ξLS. In fact, the result of this integration
obstacle to obtain the probability distribution of the is the value of the PDF of the random inputs. Thus, the
output variables.18 probability of the output constraint can be obtained by
Projection of the Feasible Region. In this work, computing the value of the PDF. A numerical integra-
we avoid direct computing of the output probability tion of (10) is required when taking a joint distribution
distribution. Instead, we derive an equivalent repre- function into account. Moreover, because of the model
sentation of the probability by mapping it to an integra- complexity, an explicit expression of (9) is usually not
tion in a region of the random inputs. Consider the available. Therefore, a Newton-Raphson step has to be
confined feasible region that will be formed by the used for computing the bound value ξLS with given ySP i ,
nonlinear projection from the region of the uncertain u, and ξ1, ..., ξS-1 for each integrated subspace. These
variables at some given u. For a practical engineering will be discussed in more detail in the next section.
problem, it is realistic that one can find a monotone Illustrative Example. The above approach is il-
relation between an output variable yi ∈ Yi and one of lustrated with the following simple problem with two
the uncertain input variables ξS ∈ ΞS, where ΞS is a uncertain variables that have a joint normal distribu-
subspace of Ξ. When this monotone relation is denoted tion. The data describing the stochastic properties of the
as yi ) F(ξS), the mapping between yi and ξS can be random inputs are given in Table 1. The output variable
schematically depicted as in Figure 2. has a nonlinear relation with these variables:
3624 Ind. Eng. Chem. Res., Vol. 41, No. 15, 2002
∫-∞∞‚‚‚∫-∞∞∫ ξLS
ξlS
F(ξ1,...,ξS-1,ξS) dξS dξS-1 ... dξ1 (13)
In this section, we derive a numerical approach to the
multivariate integration by using collocation on finite
elements for computing values of the chance constraints
The whole computational strategy proposed for solving (13). This computation depends decisively on the dis-
the nonlinear chance-constrained optimization problem tribution of the uncertain variables concerned. In pro-
is a sequential NLP approach that can be depicted in cess engineering, many uncertain variables may be
Figure 4. SQP is chosen for searching for the decision approximately treated as normally distributed. Accord-
Ind. Eng. Chem. Res., Vol. 41, No. 15, 2002 3625
Figure 3. Mapping between the input (left) and output (right) variables in an illustrative example.
where µ and Σ are known expected values and the ΦS-1(z(1) (1) (1)
2 ,..., zS ,Σ̂ ) )
covariance matrix of these stochastic variables, which
∫-∞z ‚‚‚∫-∞z φ̂S-1(ξ̂2,..., ξ̂S) dξ̂S ... dξ̂2
(1) (1)
[] [ ]
have the following form: 2 S
(24)
µ1 σ12 σ1σ2r12 ‚‚‚ σ1σSr1S Continuing this procedure for S - 2 steps, we arrive at
µ σσr σ22 ‚‚‚ σ2σSr2S the two-dimensional integration
µ) 2 Σ ) 1 2 12 (19)
‚‚‚ ‚‚‚ ‚‚‚ ‚‚‚ ‚‚‚
(S-2) (S-2) (S-2)
µS σ1σSr1S σ2σSr2S ‚‚‚ σS2 Φ2(zS-1 ,zS ,Σ̂ ))
∫ ∫
(S-2)
zS-1 z(S-2)
-∞ -∞
S
φ̂2(ξ̂S-1,ξ̂S) dξ̂S dξ̂S-1 (25)
where σi is the standard deviation of each individual
random variable and ri,j ∈ (-1, 1) is the correlation
coefficient between ξi and ξj (i, j ) 1, ..., S). ξ can be where Σ̂(S-2) is the 2 × 2 correlation matrix. Similar to
easily transformed through a linear transformation to (21), the right-hand side of (25) can be written as
3626 Ind. Eng. Chem. Res., Vol. 41, No. 15, 2002
[ z(S-2)
]
(S-2)
- r1,2 ξ̂S-1
∫-∞z
(S-2) S
S-1
Φ1 φ1(ξ̂S-1) dξ̂S-1 (26)
x (S-2) 2
1 - (r1,2 )
where Φ1 is the standard probability function of a Figure 5. Flowsheet of the reactor network.
normally distributed variable. Because an analytical
solution of (26) is not available, a numerical integration computed from (31) upward to (27), also by using the
is required. Collocation on finite elements has been collocation method.
regarded as an efficient method;21 thus, we use it for
the numerical integration of (26). In the collocation 5. Application to Process Engineering Problems
(S-2)
framework, the integrated interval (-∞, zS-1 ] will be Reactor Network Design. The first example is a
discretized into subintervals. It should be noted that for reactor network design problem, as shown in Figure 5.
the standard normal distribution the domain [-3, This example is taken from Rooy and Sahinidis,22 and
(S-2)
zS-1 ] can be used. The method of collocation on finite it is extended for the purpose of optimal design under
elements is briefly presented in appendix A. uncertainty. The design problem is defined as the cost
Based on (20)-(26), a nested computational approach minimization under the product specification. It is
can be developed. From the collocation method, the assumed that the uncertainties are from the kinetic
values of Φ2 on the collocation points of the random parameters (the activation energy and the frequency
variable will be calculated. These values will be used factor in the Arrhenius equation), while the decision
for computing the values of Φ3, and the procedure variables are the volumes of both reactors. The chance-
continues upward until the values of ΦS are computed. constrained optimization problem is formulated as fol-
The choice of the number of intervals and the number lows:
of collocation points depends on the compromise be-
tween the accuracy and computation expense. More min f ) xV1 + xV2
intervals and more collocation points will lead to more
accurate results, but the computation time will be s.t. CA1 + k1CA1V1 ) 1
increased. Simulation results show that three intervals
with three-point collocation or two intervals with five- CA2 - CA1 + k2CA2V2 ) 0
point collocation can reach a probability accuracy with
an error of less than 1%. CB1 + CA1 + k3CB1V1 ) 1
Computing the Gradients. For the case of normal
distribution, the gradients of the probability to the CB2 - CB1 + CA2 - CA1 + k4CB2V2 ) 0
decision variables u can be computed by using the same
approach as that described above. We note that only the k1 ) k10e-E1/RT1
last bound value of the integral in (20) has something
to do with u [see also (9)]. From (21), we have the
gradient vector: k2 ) k10e-E1/RT2
[ ]
value deviation matrix
E1 6665.948 200 1 0.5 0.3 0.2
0.5 1 0.5 0.1
E2 7965.248 240
0.3 0.5 1 0.3
k10 0.715 0.0215 0.2 0.1 0.3 1
k20 0.182 0.0055
[ ]
0.52 0.90 3.808 3.452 3.795 3.416 3.899 3.706 1.0 0.2 10 50.85
0.52 0.95 3.854 4.001 3.963 1.2 5
value deviation matrix
0.54 0.90 4.474 3.910 4.908 4.168 4.331 4.019 1.1 0.2 10 50.33 ηR 0.8 0.016 1 0.5 0.3 0.1
0.54 0.95 4.701 5.439 4.501 2.0 5
0.5 1 0.3 0.1
ηS 0.8 0.016
0.3 0.3 1 0.3
L
k20 } according to (7) and (9). Thus, we can use k20 L
)
-1 SP F (L/h) 37 1.11 0.1 0.1 0.3 1
F (E1,E2,k10,CB2,V1,V2) according to (10) as the upper
xF 0.2 0.008
bound for the random variable k20 in the numerical
integration. is considered (Figure 6). The column has a diameter of
The three-point collocation and four intervals for each 100 mm and 20 bubble-cap trays with a central down-
random variable are used in the multivariate integra- comer. The column is operated under atmospheric
tion, with which the inaccuracy of the probability pressure. The feed to the column is assumed to come
computation will be less than 10-4. Table 3 shows the from upstream plants and thus the feed flow rate F and
results of the stochastic optimization compared with composition xf are considered as external uncertain
those of the deterministic optimization. In the deter- variables. The tray efficiency in the stripping section
ministic optimization, the random parameters are set ηS and in the rectifying section ηR usually cannot be
constant with their expected values. In the first two determined accurately. Thus, these two parameters are
columns of Table 3, six cases (three product specifica- also considered to be uncertain variables.
tions and two probability levels) are specified. It can be A rigorous model composed of component and energy
seen that the volumes of both reactors should be balances, vapor-liquid equilibria, and tray hydraulics
increased (which thus leads to higher costs), if the for each tray is used to describe the column, which leads
product specification rises. For a same product specifi- to a complicated nonlinear equation system. The model
cation, the volumes should also be increased, if a higher is validated by comparison of the simulation and
probability level is required. For this problem, the experimental results on the plant.23 The objective func-
maximum reachable probabilities for the cases of CB2 tion of the optimization problem is defined as
g 0.5, 0.52, and 0.54 are 1.0, 0.999, and 0.971, respec-
tively. It can be noted that the results are more sensitive max f ) c1D - c2L (33)
to the predefined probability level R when R approaches
the maximum reachable probability of satisfying the
where D and L are the distillate flow rate and the reflux
constraint.
rate of the column, respectively. These two variables are
In comparison to the results of the stochastic optimi-
the decision variables of the optimization problem. The
zation, the required volumes from the deterministic
physical meaning of (33) is the maximization of the top
optimization are smaller and thus the total cost is lower.
product amount, while the operating energy should be
However, the results from the deterministic optimiza-
lowered. The optimization problem is subject to the
tion cannot be applied because the product constraint
equalities of the nonlinear equation system and a
then will be very likely violated. The last two columns
probabilistic constraint which is assigned to the top
show the percentage of the total number of violations
product, namely,
(NV %) of the constraint when implementing respec-
tively the stochastic and deterministic optimization
results, which were obtained by sampling the random D } g R
P{xD g xSP (34)
parameters, according to Table 2, with 10 000 samples
by the Monte Carlo method. As expected, about 10% (in where xD and xSPD are the distillate composition and the
the case of R ) 0.9) and 5% (in the case of R ) 0.95) of purity specification, respectively. The constants in (33)
violations will result from the results of the stochastic are set to c1 ) 4 and c2 ) 1.
optimization, while implementing the results of the The values of an assumed joint normal distribution
deterministic optimization will lead to about 50% of of the random parameters are listed in Table 4. These
violations of the constraint. Moreover, it can be seen data are estimated based on experimental experience
that the difference between the CPU times (SUN Ultra on the pilot plant. It should be noted that there is a
1 Station) of the two methods is insignificant. strong positive correlation between the two tray ef-
Operation of a Distillation Column. A pilot distil- ficiencies because both of them may increase or decrease
lation column to separate a methanol-water mixture synchronously corresponding to load changes. Feed flow
3628 Ind. Eng. Chem. Res., Vol. 41, No. 15, 2002
Table 5. Stochastic (ST) and Deterministic (DT) computing the maximum probability. It is close to 1.0
Optimization Results for both selected product specifications.
D L f
xSP R ST DT ST DT ST DT 6. Conclusions
D
0.995 0.95 0.848 0.938 1.834 1.765 1.5575 1.9862 In this work a new approach to solving nonlinear
0.995 0.90 0.867 1.820 1.6473 chance-constrained optimization problems under un-
0.995 0.5 0.931 1.754 1.9683 certainty is proposed. This approach is applicable to
0.990 0.95 0.852 0.943 1.516 1.501 1.8925 2.2697
0.990 0.90 0.871 1.506 1.9780 cases of a constrained output variable in the chance
0.990 0.5 0.938 1.470 2.2806 constraint having a nonlinear relation to uncertain
variables with joint multivariate distribution, if at least
and feed composition may have correlation due to the one of those uncertain variables has a monotonic rela-
operation of the upstream plant. For example, consider- tion to the constrained output. The monotony can be
ing a reactor as the upstream plant, if the feed flow of exploited by mapping the output feasible region to the
the reactor increases randomly, the composition of the region of uncertain variables. Through this reverse
educt component in the outflow (feed flow of the column) projection, an upper bound of the uncertain variable can
of the reactor may be increased as well because of an be computed and the required probability can be com-
uncompleted reaction. puted by numerical integration. The collocation method
is used in the integration. It is well-known that it is a
It is known in the design of distillation columns that highly efficient integration method. In the proposed
the influence of feed flow is seldom considered. However, numerical integration, the same collocation points are
feed flow has a significant impact on the liquid and also used for computing the gradients. Therefore, the
vapor loads in the column. This leads to a pressure drop advantage of direct numerical integration of the PDF
of the column and thus has an influence on the product is that one can compute the gradients simultaneously
composition (especially bottom composition). Because during the multivariate integration. This approach is
changes of the feed flow are more critical to load changes applied to two process optimization problems under
than changes of the feed concentration, the correlation uncertainty.
of the feed flow to the tray efficiencies is much stronger The knowledge of a monotone relationship between
than that of the feed concentration. The standard the constrained output and one of an input uncertain
deviations are assumed to be 2% of the expected values variable is the essence of the approach. This means that
of the tray efficiencies, 3% of the feed flow rate, and 4% to carry out optimization under uncertainty by the
of the feed concentration. proposed approach, the impact of the uncertain vari-
Now we have to chose one of the uncertain variables, ables to the output variables should be analyzed at first,
which is monotone with the restricted state variable in based on the properties of the process. This can also be
the probabilistic constraint xD for computing the bound done by using process simulation.
of the region of all uncertain variables. It can be easily The solution approach itself (described in section 3)
found out that there is a positive monotony between xF is not dependent on the distribution of the random
and xD. That means there is the relation xF v w xD v. variables. If the PDF of the random variables is known
Thus, according to (7) and (8), we can transform P{xD (e.g., uniform, exponential, normal, or log-normal dis-
g xSPD } to 1 - P{xF e xF}. According to (10), xF )
L L tribution), the approach can be applied. However, for
SP
F-1(ηR,ηS,F,xD ,D,L) can be used as the upper bound numerical integration, we have only derived the method
for the random variable xF in the numerical integration. for normally distributed random variables given in
section 4. The application of the approach to processes
The stochastic optimization problem is solved for
under uncertainties with other distributions will be a
different product specifications and probability levels. challenging future work.
The results, shown in Table 5, are compared with those
We eliminate the equality constraints by expressing
of the deterministic optimization. It is hardly a surprise the state variables in terms of the decision and uncer-
to note that an increase of the product specification and tain variables in an implicit way, namely, using simula-
the probability level causes an increase of the reflux and tion of the process by the Newton method. If a monotone
a decrease of the distillate stream and thus a reduction relation exists between the constrained output variable
of the objective function. However, it is interesting to and one selected input variable, we can be sure that one
note that the objective function and the decision vari- solution point exists when the original function is
ables are much more sensitive to the probability level replaced by the reverse function (see Figure 1). If the
than to the product specification. Similar to the previous process is too complex, there may be problems in
example, the results of the deterministic optimization simulation for the mapping of the output constraint to
(based on the expected values of the uncertain param- the uncertain input region. Then plenty of work may
eters) are close to but not the same as the results be required to overcome convergence problems. Very
computed with a probability level of 0.5. often, convergence properties strongly depend on the
The required CPU time for solving this problem initial values of the state variables. In our method, the
depends strongly on the choice of initial values of the previously computed solution point is used to ensure
decision variables. One iteration step of the optimizer that the initial values are sufficiently close to the new
takes approximately 3 min with a SUN Ultra 1 station. solution point.
A skillful choice of initial values causes about seven If a joint constraint is considered, it is necessary to
iteration steps. It should be noted that to ensure a have an uncertain variable that has a monotone rela-
feasible solution for different product specifications, it tionship with all of the chance-constrained variables.
is not necessary to compute the maximum probability For a concrete process, it may be difficult to find one
with the NLP solver in this example. Instead, one such uncertain variable. This will be a challenging task
simulation step with the total reflux is sufficient for in the future work. Another future work is to extend
Ind. Eng. Chem. Res., Vol. 41, No. 15, 2002 3629
the approach to dynamic optimization problems with (2) Subrahmanyam, S.; Pekny, J. F.; Reklaitis, G. V. Design of
time-dependent random variables (i.e., stochastic pro- Batch Chemical Plants under Market Uncertainty. Ind. Eng.
cesses). In addition, to minimize the CPU time, the Chem. Res. 1994, 33, 2688.
(3) Pistikopoulos, E. N.; Ierapetritou, M. G. Novel Approach for
collocation method used for numerical integration can Optimal Process Design under Uncertainty. Comput. Chem. Eng.
be further studied with respect to the optimal number 1995, 19, 1089.
of collocation points. (4) Rooney, W. C.; Biegler, L. T. Incorporating Joint Confidence
Regions into Design under Uncertainty. Comput. Chem. Eng. 1999,
Acknowledgment 23, 1563.
(5) Diwekar, U. M.; Kalagnanam, J. R. Efficient sampling
We thank Deutsche Forschungsgemeinschaft (DFG) technique for optimization under uncertainty. AIChE J. 1997, 43,
for financial support from the program "Echtzeit-Opti- 440.
(6) Bernardo, F. P.; Pistikopoulos, E. N.; Saraiva, P. M.
mierung grosser Systeme".
Integration and Computational Issues in Stochastic Design and
Planning Optimization Problems. Ind. Eng. Chem. Res. 1999, 38,
Appendix A: Collocation on Finite Elements for 3056.
Numerical Integration (7) Torvi, H.; Herzberg, T. Estimation of Uncertainty in Dy-
namic Simulation Results. Comput. Chem. Eng. 1997, 21 (Suppl.),
The integration required has the form S181.
(8) Acevedo, J.; Pistikopoulos, E. N. Stochastic Optimization
Ω) ∫vv ω(v) dv
f
0
(A1) Based Algorithms for Process Synthesis under Uncertainty. Com-
put. Chem. Eng. 1998, 22, 647.
(9) Prékopa, A. Stochastic programming; Kluwer: Dordrecht,
namely, The Netherlands, 1995.
(10) Ahmed, S.; Sahinidis, N. V. Robust Process Planning under
dΩ
) ω(v), Ω(v0) ) 0 (A2) Uncertainty. Ind. Eng. Chem. Res. 1998, 37, 1883.
dv (11) Gupta, A.; Maranas, C. D. A Two-Stage Modeling and
Solution Framework for Multidite Midterm Planning under
The domain [v0, vf] of the variable v will be discretized Demand Uncertainty. Ind. Eng. Chem. Res. 2000, 39, 3799.
into subintervals ([vl, vl+1], l ) 1, ..., L). In each interval (12) Clay, R. L.; Grossmann, I. E. A Disaggregation Algorithm
the integration function will be approximated by for the Optimization of Stochastic Planning Models. Comput.
Lagrange functions Chem. Eng. 1997, 21, 751.
(13) Kall, P.; Wallace, S. W. Stochastic programming; Wiley:
NC New York, 1994.
Ω(v) ) ∑
j)0
Γj(v) Ωj (A3) (14) Schwarm, A. T.; Nikolaou, M. Chance-constrained Model
Predictive Control. AIChE J. 1999, 45, 1743.
(15) Li, P.; Wendt, M.; Wozny, G. Robust Model Predictive
where NC is the number of collocation points and Γ is Control under Chance Constraints. Comput. Chem. Eng. 2000, 24,
the Langrange function 829.
(16) Whiting, W. B.; Tong, T.; Reed, M. E. Effect of Uncertain-
ties in Thermodynamic Data and Model Parameters on Calculated
NC
v - v(i)
Γj(v) ) ∏ (A4) Process Performance. Ind. Eng. Chem. Res. 1993, 32, 1367.
(17) Vasquez, V. R.; Whiting, W. B. Uncertainty and Sensitivity
i)0 v(j) - v(i) Analysis of Thermodynamic Models Using Equal Probability
i*j
Sampling (EPS). Comput. Chem. Eng. 2000, 23, 1825.
Ωj and v(j) are the values of the integration and the (18) Morgan, M. G.; Henrion, M. Uncertainty; Cambridge
corresponding collocation points which are usually the University Press: Cambridge, U.K., 1990.
(19) Loeve, M. Probability Theory; Van Nostrand-Reinhold:
zeros of the shifted Legendre function,21 respectively. Princeton, NJ, 1963.
Thus, the derivatives of Ω on the collocation points can (20) Papoulis, A. Probability, Random Variables and Stochastic
be calculated as Processes; McGraw-Hill: New York, 1965.
(21) Finlayson, B. A. Nonlinear Analysis in Chemical Engineer-
| | Ωj,
dΩ NC dΓj
∑
ing; McGraw-Hill: New York, 1980.
) i ) 1, ..., NC (A5) (22) Rooy, H. S.; Sahinidis, N. V. Global Optimization of
dv vi j)0 dv v i Nonconvex NLPs and MINLPs with Applications in Process
Design. Comput. Chem. Eng. 1995, 19, 551.
From (A2) and (A5), the value of the integration on the (23) Li, P.; Wendt, M.; Arellano, H. G.; Wozny, G. Optimal
collocation points can be calculated. For the continuity, Operation of Distillation Processes under Uncertain Inflows
we use the integrated value on the last collocation point Accumulated in a Feed Tank. AIChE J. 2002, 48, 1198.
of the current interval as the initial value of the next
interval. Received for review July 31, 2001
Revised manuscript received April 10, 2002
Literature Cited Accepted May 22, 2002
(1) Halemane, K. P.; Grossmann, I. E. Optimal Process Design
under Uncertainty. AIChE J. 1993, 29, 425. IE010649S