1 This communication deals with mass customization and the association of the product configuration task with the planning of its production process while trying to minimize cost and cycle time. We consider a two steps approach that first permit to interactively (with the customer) achieve a first product configuration and first process plan (thanks to non-negotiable requirements) and then optimize both of them (with remaining negotiable requirements). The communication concerns the second optimization step. Our goal is to evaluate a recent evolutionary algorithm (EA). As both problems are considered as constraints satisfaction problems, the optimization problem is constrained. Therefore the considered EA was selected and adapted to fit the problem. The experimentations will compare the EA with a conventional branch and bound according to the problem size and the density of constraints. The hypervolume metric is used for comparison..
This paper deals with mass customization and more accurately with
aiding the two activities, product configuration and production
planning, achieved in a concurrent way. According to the
preferences of each customer, the customer requirements
(concerning either the product or its production) can be either
nonnegotiable or negotiable. This situation allows considering a
twostep process that aims to associate the two conflicting expectations,
interactivity and optimality. The first interactive step, that
sequentially processes each non-negotiable requirement,
corresponds with a first configuration and planning process that
reduces the solution space. This process is present in many
commercial web sites using configuration techniques like
automotive industry for example. Then, a second process optimizes
the solution with respect to the remaining negotiable requirements.
As the solution space can quickly become very large, the
optimization problem can become hard. Thus, this behavior is not
frequent in commercial web sites. Meanwhile some scientific
works have been published on this subject (see for example [
Deriving the definition of a specific or customized product
(through a set of properties, sub-assemblies or bill of materials,
etc…) from a generic product or a product family, while taking
into account specific customer requirements, can define product
configuration [
This concurrent process and the supporting constraint framework
present three main interests. First they allow considering constraint
that links configuration and planning in both directions (for
example: a luxury product finish requires additional manufacturing
time or a given assembly duration forbids the use of a particular
kind of component). Secondly they allow processing in any order
product and planning requirements, and therefore avoid the
traditional sequence: configure product then plan its production [
In order to illustrate the addressed problem we consider a very simple example dealing with the configuration and planning of a small plane. The constraint model is shown in figure 1. The plane is defined by two product variables: number of seats (Seats, possible values 4 or 6) and flight range (Range, possible values 600 or 900 kms). A constraint Cc1 forbids a plane with 4 seats and a range of 600 kms. The production process contains two operations: sourcing and assembling. (noted Sourc and Assem). Each operation is described by two process variables: resource and duration: for sourcing, the resource (R-Sourc, possible resources “Fast-S” and “Slow-S”) and duration (D-Sourc, possible values 2, 3, 4, 6 weeks), for assembling, the resource (R-Assem, possible resources “QuicA” and “Norm-A”) and duration (D-Assem, possible values 4, 5, 6, 7 weeks). Two constraints linking product and process variables modulate configuration and planning possibilities: one linking seats with sourcing, Cp1 (Seat, R-Sourc, D-Sourc), and a second one linking range with the assembling, Cp2 (Range, R-Assem, DAssem). The allowed combinations of each constraint are shown in the 3 tables of figure 1. Without taking constraints into account, this model shows a combinatory of 4 for the product (2x2) and 64 for the production process (2x4) x (2x4) providing a combinatory of 256 (4 x 64) for the whole problem. Considering constraints lead to 12 solutions for both product and production process. Given previous problem, various criteria can characterize a solution: on the product configuration side, performance and product cost, and on the production planning side, cycle time and process cost. In this paper we only consider cycle time and cost. The cycle time matches the ending date of the last production operation of the configured product. Cost is the sum of the product cost and process cost. We are consequently dealing with a multicriteria optimization problem. As these criteria are in conflict, it is better for decision aiding to offer the customer a set of possible compromises in the form of Pareto Front.
Figure 2 Concurrent configuration and planning model to optimize
In order to complete our example, we add a cost and cycle time
criteria as represented in figure 2. For cost, each product variable
and each process operation is associated with a cost parameter and
a relevant cost constraint: (C-Seats, Cs1), (C-Range, Cs2),
(CSourc, Cs3) and (C-Assem, cs4) detailed in the tables of figure 2.
The total cost is obtained with a numerical constraint and the cycle
time, sum of the two operation durations, is also obtained with a
numerical constraint as follow:
Total cost = C-Seats + C-Range + C-Sourc + C-Assem.
Cycle time = D-Sourc + D-Assem
The twelve previous solutions are shown on the figure 3 with the
Pareto front gathering the optimal ones. In this figure, all solutions
are present. When non-negotiable requirements are processed
during interactive configuration and planning, some of these
solutions will be removed. Once all these requirements are
processed, the identification of the Pareto front can be launched in
order to propose the customer a set of optimal solutions.
A strong specificity of this kind of problems is that the solution
space is large. It is reported in [
In the following section we characterize the optimization problem and briefly recall the optimization techniques. Then experimentation results are presented and discussed in the last section.
The problem of figure 2 is generalized as the one shown in figure 4. The optimization problem is defined by the quadruplet <V, D, C, f > where V is the set of decision variables, D the set of domains linked to the variables of V, C the set of constraints on variables of V and f the multi-valuated fitness function. Here, the aim is to minimize both cost and cycle time. The set V gathers: the product descriptive variables and the resource variables. The set C gathers constraints (Cc and Cp). Cost variables and operation durations are deduced from the variables of the set V thanks to the remaining constraints. Experimentations will consider different problem sizes: different numbers of product variables, different number of production operations and different number of possible values for these variables. Different constraint densities (percentage of excluded combinations of values) will be also considered. 2.2
The proposed evolutionary algorithm is based on SPEA2 [
For initialization, crossover and mutation operators, each time an
individual is created or modified, every gene (decision variable of
V) is randomly instantiated into its current domain. To avoid the
generation of unfeasible individuals, the domain of every
remaining gene is updated by constraint filtering. As filtering is not
full proof, inconsistent individuals can be generated. In this case a
limited backtrack process is launched to solve the problem. For full
details please see [
The key idea of the Branch and Bound algorithm is to explore a search tree but using a cutting procedure that stops exploration of a branch when a better branch has already been found. The first tool is a splitting procedure that corresponds to the selection of one variable of the problem and to the instantiation of this variable with each possible value. The second tool is a node-bound evaluation procedure. The filtering process is used to achieve this task with a partial instantiation and is able to evaluate if the partial instantiation is consistent with the constraints of the problem, and, if this is the case, to provide the lower bound of each criterion cycle time and cost. When the search reaches a leaf of the search tree, or complete instantiation, the filtering system gives the exact evaluation of the solution. Thus, the values of leaf solutions can be used to compute the current Pareto front and then to cut remaining unexplored branches that are dominated by any aspect of the Pareto front solution (e.g. the upper bounds of the leaf solution dominate the minimal bounds of the branch to cut). 3
The optimization algorithms were implemented in C++ programming language and interacted with a filtering system coded in Perl language. All tests were done using a laptop computer powered by an Intel core i5 CPU (2.27 Ghz, only one CPU core is used) and using 2.8 Go of ram. These tests compared the behavior of our constrained EA algorithm with the exact branch-and-bound algorithm. 3.1
constraint densities problem size and An initial first model, named "full model" is considered. It can be consulted and interactively used at http://cofiade.enstimac.fr/cgibin/cofiade.pl select model ‘Aircraft-CSP-EA-10’. It gathers five product variables with a domain size between 4 and 6, six production operations with a number of possible resources between 3 and 25. Without constraints consideration, the solution space of the product model is 5,184, and the planning model is 96,000. The size of the global problem model is 497,664,000. A second model, named “small model”, has been derived from the previous one with the suppression of a high combinatory task and a reduction of one domain size. This reduces the planning problem size to 12,000 and global model 6,220,800.
In order to evaluate the impact of constraints density, two versions
of each model were produced: one with a "weak density" of
constraints (20% of possible combinations are excluded in each
constraint Cc and Cp) and the other with a "high density" of
constraints (50% excluded). These values are frequently met in
industrial configuration situations. This provides four models
characteristics in table 1.
For the small models, evolutionary settings are tuned to: population
size: 50; archive size: 40; Pmut: 0.4; Pcross: 0.8. The ending criterion
used is a time limit of half an hour. For the full models, we adapt
settings for a wider search: population size: 150; archive size: 100;
Pmut: 0.4; Pcross: 0.8. The ending criterion used is a time required by
the BB algorithm. In order to analyze the two optimization
approaches, we compare the hypervolume evolution during
optimization process. Hypervolum metric has been defined in [
Results are presented in figure 6 where EA curves are average results for 30 executions. Both algorithms start with a lapse of time where performance is null. For the BB algorithm, this corresponds to the time needed to reach a first leaf on the search tree, while for the EA; it corresponds to the time consumed to constitute the initial population.
For the small models (first two curves), the BB algorithm reaches the optimal Pareto front much faster compared with EA performance. On the other hand, the EA is logically better than the BB algorithm on the full model. For example, on the lowconstrained model, the BB algorithm took 20 times longer to reach a good set of solutions (less than 0.5% of the optimal hypervolume).
The impact of constraints density could also be discussed. As it can be seen, the BB algorithm performance is improved when the density of constraints is high. This is because the filtering allows more branches to be cut on the search tree, in such way that the algorithm reaches leaf solutions and, consequently, optimal solutions more quickly. The EA performance moves in the opposite way. The more the model is constrained, the more the random crossover operation will have to backtrack to find feasible solutions, and thus the time needed by the algorithm will be consequent.
Figure 6 First experimentation results In order to try to identify the problem size where EA is more suitable than BB, we have modified the low constrained model as follows. We consider now a model gathering six product variables and six production operations with three possible values for each, and sequentially add either a product variable and or an operation. The range of study is between 12 and 16 decision variables with three possible values for each. Relevant solution spaces without constraint vary between 1.6*106 and 43*106.
The results are shown in the left part figure 7. The vertical axis corresponds with the computation time and the horizontal one with the number of decision variables. For BB curves, it shows the time to reach the optimal solution. For the EA curve it shows the time required for nine EA runs over ten to reach the optimal solution. Order of magnitude are close for both around 13 or 14 variables corresponding with a solution space around 2*106 to 5*106 comparable with our previous small model size.
As we already mention, industrial models are frequently larger than that. We therefore try our EA approach with a low constrained model with 30 variables and a solution space around 1016. The stopping criterion is "2 hours without improvement". The right part of Figure 7 shows that the optimization process has stopped after 48 hours. It can be noticed that 90% of the final score was obtained after 3 hours and 99% in 10 hours. This allows underlining the good performance of our approach when facing large low constrained problems. Of course the idea is to use BB, if the first interactive configuration step has led to a rather small problem, less than 13 or 14 variables in our case, and EA otherwise.
Finally we also try to break optimization in two steps. The idea is: (i) compute quickly a low quality Pareto, (ii) select the area that interest the customer (iii) compute a Pareto on the restricted area. The restricted area is obtained by constraining the two criteria total cost and cycle time (or interesting area) and filtering these reductions on the whole problem. The search space is greatly reduced and the second optimization much faster. This is shown in figure 8 where the left part shows the single step process with 10 and 60 minutes Pareto and the right part shows the restricted area with the two previous curve and the one corresponding with a 10 minutes Pareto launched on the restricted area. It shows that the sequence of two optimization steps of 10 minutes provide a result equivalent to a 60 minutes optimization process.
Figure 8 Experimentations with a two steps optimization process
The goal of this communication was to propose a first evaluation of an adapted evolutionary algorithm that deals with concurrent product configuration and production planning. The problem was recalled and the two optimization approaches (Evolutionary algorithm and branch and bound) where briefly presented. Various experimentations have been presented. A first result is that: (i) the proposed EA works fine when the size of the problem gets large compare to the BB, (ii) when problem tends to be more constrained the tendency goes to the opposite. When problem is low constrained (90% of excluded solutions) with 13-14 decision variables with 3 values each, they perform equally. When the problem gets larger, BB cannot be considered and EA can provide good quality results for the same problem with up to 30 variables (around 1016 solutions - 90% rejected). Finally some ideas about a two steps optimization process have shown that the proposed approach is quite promising for large problems. These are first experimentation results and we are now working on comparing our proposed EA with some penalty function approaches.
We would like to thank the referees for their comments, which helped improve this paper considerably.