## Derivative-free optimization methods for handling fixed costs in optimal groundwater remediation design (2006)

Venue: | in Proc. of the CMWR XVI - Computational Methods in Water Resources |

Citations: | 7 - 4 self |

### BibTeX

@INPROCEEDINGS{Hemker06derivative-freeoptimization,

author = {T. Hemker and K. R. Fowler and O. Von Stryk},

title = {Derivative-free optimization methods for handling fixed costs in optimal groundwater remediation design},

booktitle = {in Proc. of the CMWR XVI - Computational Methods in Water Resources},

year = {2006}

}

### OpenURL

### Abstract

We consider a hydraulic capture application for water resources management that includes a fixed installation cost in addition to operating costs. The result is a simulationbased, nonlinear, mixed-integer optimization problem. The motivation is that our preliminary studies have shown that convergence to an unsatisfactory, local minimum with many wells operating at low pumping rates is common when the fixed cost is ignored. Such optimization tasks are not unique to subsurface management, rather efficient simulation-based methods are needed in the whole field of computational engineering. All the approaches used below do not need the gradient of the objective function, only function values for minimization. In one approach, we bypass including the number of wells as a decision variable by defining an inactive-well threshold. In another approach, we use penalty coefficients proposed in the literature to transform the discontinuous problem into a continuous one. For the two above formulations, we use the implicit filtering algorithm. In the third approach, we introduce a mixed-integer problem formulation and use an iterative stochastic modeling technique to build surrogate functions that approximate the objective function. With this new procedure the use of a branch-and-bound technique becomes possible to solve the mixed-integer problem in contrast to methods working directly on the simulation results, which impedes relaxation of integer variables. We present promising numerical results on the benchmarking problem and point the way towards improvement and future work.

### Citations

1544 |
Practical Optimization
- Gill, Murray, et al.
- 1981
(Show Context)
Citation Context ... (s(j) ,Q (j) ,x (j) ,y (j) ), for j = 1,...,N, and k = 1,..,M. These NLP-subproblems at the different knots and leaves of the BB-tree are be solved via sequential-quadratic-programming (SQP) methods =-=[10]-=-, because numerical noise is avoided by working on the surrogate functions. In our approach we use SNOPT for Matlab [9], which returns the best candidate found, and its merit function value to the BB ... |

70 |
A Modular Three Dimensional Finite Difference Groundwater Flow Model’. U.S. Geological Survey Techniques of Water Resources Investigations
- McDonald, Harbaugh
- 1988
(Show Context)
Citation Context ...m] for well i, also vary with the decision variables and obtaining their values at each iteration requires a solution to the partial differential equation that models saturated flow. We use Modflow96 =-=[17]-=- for the saturated flow simulation. We impose the following constraints; −0.0064 ≤ Qi ≤ 0.0064[m 3 /s],i = 1,...,n (2) 30 ≥ hi ≥ 10[m],i = 1,...,n (3) QT = n∑ Qi ≥ −0.032[m 3 /s]. (4) i=1 To contain t... |

61 | E.: 2002, ‘Review of Nonlinear Mixed-Integer and Disjunctive Programming techniques
- Grossmann
(Show Context)
Citation Context ...resent all possible zero-one combinations for s. All branches in the BB-tree with higher function values than prior found feasible solutions are discarded from further evaluations. See also Grossmann =-=[12]-=- who characterizes such techniques based on the decomposing the optimization problem, or the textbook of Floudas [6] for a broader introduction on these methods. The problem considered here is challen... |

54 | An implicit filtering algorithm for optimization of functions with many local minima
- Gilmore, Kelley
- 1995
(Show Context)
Citation Context ...that minimizes ˜ J = ˜ J c + ˜ J o . 4. OPTIMIZATION APPROACHES 4.1. IMPLICIT FILTERING. Implicit filtering (IF) is a projected quasi-Newton method that uses a sequence of finite difference gradients =-=[11]-=-. The difference increment is reduced as the optimization progresses to take advantage of the fast convergence of quasiNewton methods near a local minimum. Because (IF) relies on finite difference gra... |

44 |
A Matlab Kriging Toolbox
- Lophaven, Nielsen, et al.
(Show Context)
Citation Context ...etween the trend and the interpolation points. For more details in theory of computer experiments, compare to [14]. The DACE-models here are evaluated by the Matlab Kriging Toolbox of Lophaven et.al. =-=[15]-=-. This consideration yields a surrogate objective function ˆ J = ˜ Jc + ˆ Jo , which no longer depends on simulation evaluations. Under the assumption, that ˆ J reflects the major characteristics of ˜... |

43 | 2002, ‘Flexibility and Efficiency Enhancements for Constrained Global Design Optimization with Kriging Approximations
- Sasena
(Show Context)
Citation Context ...is points. The minimizer of the surrogate function ˆ J is added to the set of basis points for the DACE-model in the next iteration step. Such sequential procedures are discussed in greater detail in =-=[23]-=-. However, in contrast to the update strategies discussed therein, we use convergence to a previously determined system design as criteria to stop minimization on ˆ J and to search for a new simulatio... |

42 |
Deterministic Global optimization Theory, Methods and Applications
- Floudas
- 1999
(Show Context)
Citation Context ...ound feasible solutions are discarded from further evaluations. See also Grossmann [12] who characterizes such techniques based on the decomposing the optimization problem, or the textbook of Floudas =-=[6]-=- for a broader introduction on these methods. The problem considered here is challenging in that the operating cost depends on the results of the subsurface flow simulation, so that ˜ J is not relaxab... |

13 | IFFCO: Implicit Filtering for Constrained Optimization, Version 2
- Choi, Gilmore, et al.
- 1999
(Show Context)
Citation Context ...ifference gradients,CMWR XVI 5 only function values are needed to guide the minimization. For this work, we use a FORTRAN implementation called IFFCO, with the symmetric rank one quasi-Newton update =-=[5]-=-. We used the default optimization parameter settings. There are several convergence theorems for implicit filtering, which was particularly designed for the optimization of noisy functions, and indee... |

7 | 2003, ‘Mixed-Integer Nonlinear Programming
- Bussieck, Pruessner
(Show Context)
Citation Context ...n the objective is described analytically, because the process combines difficulties from both continuous and discrete optimization. An overview of optimization methods for these problems is given in =-=[4]-=-. In our approach we use a classical branch-and-bound (BB) method to guaranty the integrality constraints on the switching vector s for new candidate system designs (s,Q,x,y). The BB starts with relax... |

7 | Solution of a well-field design problem with implicit filtering - Fowler, Miller, et al. - 2004 |

7 |
User’s guide for SNOPT 7.1: a fortran package for large-scale nonlinear programming
- Gill, Murray, et al.
- 2006
(Show Context)
Citation Context ...of the BB-tree are be solved via sequential-quadratic-programming (SQP) methods [10], because numerical noise is avoided by working on the surrogate functions. In our approach we use SNOPT for Matlab =-=[9]-=-, which returns the best candidate found, and its merit function value to the BB as the result in a knot. The basis points for the DACEmodel are not determined a priori by a space filling set of syste... |

5 | Solution of a groundwater control problem with implicit filtering
- Batterman, Gablonsky, et al.
- 2000
(Show Context)
Citation Context ...gence theorems for implicit filtering, which was particularly designed for the optimization of noisy functions, and indeed IFFCO has been successfully applied to other groundwater management problems =-=[7, 8, 3]-=-. 4.2. ITERATIVE UPDATED SURROGATE-FUNCTIONS. Solving mixed-integer nonlinear problems is a challenging task, even when the objective is described analytically, because the process combines difficulti... |

5 |
Designs for computer experiments, Technometrics 31(1
- Sacks, Schiller, et al.
- 1989
(Show Context)
Citation Context ...ect to s and as described above, and gradient information is not available. To avoid these problems, we apply an extension of the classical design and analysis of computer experiments (DACE) approach =-=[22]-=- described in [13]. Here, the underlying process with real-valued and integervalued variables is approximated to build an analytic surrogate function with real-valued variables only. Under the assumpt... |

4 |
2006b, ‘Mixed-integer simulation-based optimization for a superconductive magnet design
- Hemker, Glocker, et al.
(Show Context)
Citation Context ...scribed above, and gradient information is not available. To avoid these problems, we apply an extension of the classical design and analysis of computer experiments (DACE) approach [22] described in =-=[13]-=-. Here, the underlying process with real-valued and integervalued variables is approximated to build an analytic surrogate function with real-valued variables only. Under the assumption of a real-valu... |

3 |
Optimal Design of Flow in Groundwater Systems
- Ahfeld, Mulligan
- 2000
(Show Context)
Citation Context ...future work. 1. INTRODUCTION AND MOTIVATION A hydraulic capture (HC) problem involves the placement of wells to alter the direction of groundwater flow and halt the migration of the contaminant plume =-=[2]-=-. Subsurface simulation is needed to understand the response of the aquifer and predict the fate of the plume. Optimization techniques work in conjunction with the simulators to determine the optimal ... |

3 | Which groundwater remediation objective is better: A realistic one or a simple one - Ren, Minsker |

3 |
MT3DMS: A Modular 3-D Multispecies Transport Model for Simulation of Advection
- Zheng, Wang
- 1998
(Show Context)
Citation Context ...ted from a finite source for five years prior to the remediation period with a constant concentration of 1kg/m 3 located at [(200, 225), (475, 525), (h,h−2)][m] in the physical domain. We use MT3DMS, =-=[24]-=- to generate the plume and check for containment, and use the 5 × 10 −5 contour line as the plume boundary. All methods used an initial well design with two injection wells and two extraction wells se... |

2 |
A hydraulic capture application for optimal remediation design
- Fowler, Kees, et al.
- 2004
(Show Context)
Citation Context ...tion proposed in the literature specifically for benchmarking [20]. This HC problem has been the focus of studies comparing constraint formulations, subsurface simulators, and optimization approaches =-=[16, 7]-=-. In this study we compare three approaches to handle the installation cost. In the first, the objective function remains discontinuous and we use an inactive-well threshold to remove wells from a sys... |

2 |
Pump-and-treat optimization using analytic element method flow models.” Adv
- Matott, Rabideau, et al.
- 2006
(Show Context)
Citation Context ...tion proposed in the literature specifically for benchmarking [20]. This HC problem has been the focus of studies comparing constraint formulations, subsurface simulators, and optimization approaches =-=[16, 7]-=-. In this study we compare three approaches to handle the installation cost. In the first, the objective function remains discontinuous and we use an inactive-well threshold to remove wells from a sys... |

2 |
Approximate mixed integer nonlinear programming methods for optimal aquifer remdiation design, Water Resources Research 31
- McKinney, Lin
- 1995
(Show Context)
Citation Context ...standing of optimization software. It is becoming more accepted that fixed installations costs included with operating costs are needed when remediation time horizons are short, say five or ten years =-=[18, 21]-=-. Using the cost data in [20], it costs roughly $20,000 to install an extraction well and $1,000 to operate the well for a year. Ignoring installation costs, and starting with a large set of “candidat... |

2 |
Electronic Supplement to ”Optimal Design for Problems Involving Flow and Transport
- Meyer, Kelley, et al.
- 2002
(Show Context)
Citation Context ...f wells, n ≤ N, the pumping rates {Qi} n i=1[m 3 /s], and the {(xi,yi)} n i=1 locations. The objective function is the sum of the installation (capital) cost J c and the operational cost J o given by =-=[20, 19]-=-: J = + n∑ i=1 c0d b0 i ∑ + Qi<0.0 c1|Q m i | b1 (zgs − h min ) b2 } {{ } Jc ∫ ( tf ∑ c2Qi(hi − zgs) + ∑ c3Qi 0 } i,Qi<0.0 {{ i,Qi>0.0 } Jo ) dt. (1) In Jc , the first term accounts for drilling and i... |

1 |
Optimal ground-water remediation methods applied to a superfund site: from formulation to implementation, Ground Water 33
- Ahfeld, Pinder
- 1995
(Show Context)
Citation Context ...a suboptimal final system design, with many wells operating at low pumping rates. Typically, low-pumping wells are then consolidated and optimization is re-run with the smaller set of candidate wells =-=[1]-=-. An approach that selects the appropriate number of wells in the course of the optimization is much more attractive. In this work we focus on formulations that include fixed installation costs as wel... |

1 |
Computer experiments, Handbook of Statistics 13
- Koehler, Owen
- 1996
(Show Context)
Citation Context ...nce controlled by a correlation function R, and a variance σ2 . Z models the lack-of-fit between the trend and the interpolation points. For more details in theory of computer experiments, compare to =-=[14]-=-. The DACE-models here are evaluated by the Matlab Kriging Toolbox of Lophaven et.al. [15]. This consideration yields a surrogate objective function ˆ J = ˜ Jc + ˆ Jo , which no longer depends on simu... |

1 |
design for problems involving flow and transport in saturated porous media
- Optimal
- 2002
(Show Context)
Citation Context .... It is becoming more accepted that fixed installations costs included with operating costs are needed when remediation time horizons are short, say five or ten years [18, 21]. Using the cost data in =-=[20]-=-, it costs roughly $20,000 to install an extraction well and $1,000 to operate the well for a year. Ignoring installation costs, and starting with a large set of “candidate wells”, can result in a sub... |