36
2 Simulation-Based Optimization
on numeric function evaluations and can therefore be easily used in combination
with simulation, even though they do not guarantee to produce globally optimal
solutions [Go2015, p. 90].
While AutoStat, Extend Optimizer, RISKOptimizer, and SimRunner2 use evolution strategies and/or genetic algorithms for their search procedures, OptQuest is
based on a combination of scatter search, tabu search as well as neural networks,
and WITNESS Optimizer uses simulated annealing and tabu search [La2007,
p. 660; FGA2005, p. 88]. However, the literature on the performance of those optimization packages is rather poor [AGA2016, p. 1]. The most discussed optimizers
are OptQuest and WITNESS Optimizer 18 .
OptQuest uses scatter search as its primary search procedure and combines
it with tabu search as well as neural networks. Scatter search can be applied
to problems with continuous and discrete variables, as well as on one or multiple objective optimization. Scatter search shows high-quality outcomes for hard
combinatorial optimization problems and supports the use of additional heuristics to check selected reference points for improved solutions [La2011, pp. 4–5;
MCP2018, p. 717]. The algorithm selects the best solutions as a starting point
for the next application until a defined number of iterations or one of the three
stopping rules is reached. As described by Law, scatter search stops the search
process when a user-specified number of configurations is reached, when there is
no improvement in the objective function (Automatic Stop), or when the number of non-improving configurations makes up five percent of the user-specified
number of configurations [La2007, p. 661]. Tabu search in forms of a modified
neighborhood search is included in OptQuest to keep track of relevant solutions
in tabu lists and to prohibit the reinvestigation of already evaluated solutions to
reach a global optimum [Am+2016, p. 366; Go2015, p. 95]. Neural networks
are used as a prediction model to shorten the search process by avoiding the
evaluation for reference points with a predictable low-quality value [Es+2011,
p. 2364]. OptQuest allows linear and nonlinear constraints (such as space restrictions, workforce allocations or budget limits) on decision variables as well as on
output variables and the dependence on a variance estimate for the number of replications for a particular configuration. This provides a higher statistical guarantee
that the best simulated system configuration will be returned [La2007, p. 661].
18 Following, OptQuest and WITNESS Optimizer as well as the used search procedures are
described very briefly. The interested reader is referred to Fu, Glover and April [FGA2005],
Glover, Laguna and Martí [GLM2000], Law [La2007, pp. 659–666], Hindelsberger and
Vidal [HV2000], Eskandari et al. [Es+2011], and Gosavi [Go2015, pp. 71–120 and
pp. 343–348] for further reading on OptQuest and WITNESS Optimizer as well as on scatter
search, tabu search, and simulated annealing.
2 Simulation-Based Optimization
on numeric function evaluations and can therefore be easily used in combination
with simulation, even though they do not guarantee to produce globally optimal
solutions [Go2015, p. 90].
While AutoStat, Extend Optimizer, RISKOptimizer, and SimRunner2 use evolution strategies and/or genetic algorithms for their search procedures, OptQuest is
based on a combination of scatter search, tabu search as well as neural networks,
and WITNESS Optimizer uses simulated annealing and tabu search [La2007,
p. 660; FGA2005, p. 88]. However, the literature on the performance of those optimization packages is rather poor [AGA2016, p. 1]. The most discussed optimizers
are OptQuest and WITNESS Optimizer 18 .
OptQuest uses scatter search as its primary search procedure and combines
it with tabu search as well as neural networks. Scatter search can be applied
to problems with continuous and discrete variables, as well as on one or multiple objective optimization. Scatter search shows high-quality outcomes for hard
combinatorial optimization problems and supports the use of additional heuristics to check selected reference points for improved solutions [La2011, pp. 4–5;
MCP2018, p. 717]. The algorithm selects the best solutions as a starting point
for the next application until a defined number of iterations or one of the three
stopping rules is reached. As described by Law, scatter search stops the search
process when a user-specified number of configurations is reached, when there is
no improvement in the objective function (Automatic Stop), or when the number of non-improving configurations makes up five percent of the user-specified
number of configurations [La2007, p. 661]. Tabu search in forms of a modified
neighborhood search is included in OptQuest to keep track of relevant solutions
in tabu lists and to prohibit the reinvestigation of already evaluated solutions to
reach a global optimum [Am+2016, p. 366; Go2015, p. 95]. Neural networks
are used as a prediction model to shorten the search process by avoiding the
evaluation for reference points with a predictable low-quality value [Es+2011,
p. 2364]. OptQuest allows linear and nonlinear constraints (such as space restrictions, workforce allocations or budget limits) on decision variables as well as on
output variables and the dependence on a variance estimate for the number of replications for a particular configuration. This provides a higher statistical guarantee
that the best simulated system configuration will be returned [La2007, p. 661].
18 Following, OptQuest and WITNESS Optimizer as well as the used search procedures are
described very briefly. The interested reader is referred to Fu, Glover and April [FGA2005],
Glover, Laguna and Martí [GLM2000], Law [La2007, pp. 659–666], Hindelsberger and
Vidal [HV2000], Eskandari et al. [Es+2011], and Gosavi [Go2015, pp. 71–120 and
pp. 343–348] for further reading on OptQuest and WITNESS Optimizer as well as on scatter
search, tabu search, and simulated annealing.