Professional Documents
Culture Documents
Table of contents:
Chapter I - General information......................................................................................... 2
I.A. Introduction ................................................................................................................. 2
I.B. Applications within the forging process ...................................................................... 2
I.C. Principle and definitions.............................................................................................. 3
Chapter II - Description of the algorithm .......................................................................... 4
II.A. Mono-objective algorithm ......................................................................................... 4
II.B. Evolutionary algorithm .............................................................................................. 4
II.C. Notion of the kriging metamodel ............................................................................... 5
II.D. Operation of the genetic algorithm ............................................................................ 7
II.E. Initialisation of the population ................................................................................. 10
II.F. Computation of a generation .................................................................................... 10
Chapter III - Optimisation strategies ............................................................................... 13
III.A. Initialisation by experimental design ..................................................................... 13
III.B. Discrete optimisation .............................................................................................. 14
Chapter IV - Example: optimisation of a billet volume .................................................. 15
Index: ................................................................................................................................... 17
Bibliography: ....................................................................................................................... 18
I.A. Introduction
Automatic optimisation is a new approach to problem resolution and process enhancement in
material forming. It automates long and repetitive data setup and parameter selection.
Formely, the user would launch a series of computations, progressively modifying data setup
according to results obtained for each simulation. With the automatic optimisation feature, the
user directly submits their optimisation problem to the software by choosing the criteria to be
improved, the constraints to be satisfied, and the variables to be modified. The application
will automatically launch a series of simulations with a judicious choice of parameters and the
goal of finding an optimal solution. The work required of the user is simply to structurally
formalise the objectives, the imposed constraints, and the parameters that can vary. In this
document, "minimisables" are the criteria to improve, "constraints" the constraints to satisfy,
and "parametered actions" the process variables to modify.
A first example is to minimise the weight/volume of the blocker in a forging context. The
constraints are: die completely filled and no folds. The parameters that can vary are: billet
diameter and height. The user can choose the amount and types of parameter they want to
vary in order to optimise the process. They can also choose to vary billet diameter by
imposing a series of defined values that correspond to their supplier's diameters for example.
In the optimisation example provided in Figure 1, a 1.4 kg savings of material per forged part
was accomplished. This gain also reduced the maximum tonnage required to forge the part
and, as a result, made it possible to use a smaller size press. Figure 1 draws attention to the
fact that the flash of the part has shrunk while maintaining the same filling rate.
A second example could be to minimise part temperature. The constraint is the maximum
force of the press. The parameters that can vary are the lubrication conditions and press speed.
To resolve this type of problem, a multi-objective algorithm is generally used. This algorithm
will provide a set of optimal solutions to form what we call a "Pareto Frontier". Depending on
the priority given to one objective or the other, it's a different point of the Pareto Frontier that
will be selected and become the optimal solution to the problem.
To find a solution to this type of problem, you can also use several minimisables with a
penalised mono-objective algorithm and consequently lean towards a pseudo multi-objective
algorithm that will provide a unique optimal solution. To do this, the cost function will be
obtained by a rated sum of the different minimisables chosen. This is the type of algorithm we
use in our software.
Evaluation of the
individual
Optimisation algorithm
FEM solver
Population update
Using a metamodel greatly reduces this computation cost [Ejday 2011, Roux 2011]. The
principle is to replace the direct evaluation of an individual with an estimation using a
simplified model (the metamodel) of the problem to be solved. This metamodel is constructed
using the database of evaluated individuals. It can therefore not be constructed or used until
after the complete evaluation of the first generation of individuals. With each new generation,
the metamodel becomes more precise. This metamodel is used to get a first cost function
approximation for an individual. Based on this estimation and the error estimation given to us
by the metamodel, the individual can be selected for a complete finite element evaluation
(Figure 4). This method avoids having to compute a very large number of individuals in order
to search the space. Computed individuals are relevantly chosen to both improve the
metamodel and to converge towards the optimum.
Improved
metamodel
Estimation of the
individual FEM solver
There are several types of metamodel in the literature, such as neural networks, radial basis
functions, and kriging, for example. The last method is the one used in our case. The objective
of kriging is to provide a cost function approximation for the entire domain defined by the
limits of the optimisation parameters. Each optimisation parameter represents a dimension of
the optimisation space. Kriging was originally developed empirically for geostatistics by D.
Krige in the 50s, then mathematically formalised in the 60s by G. Matheron, mathematician at
the École des Mines de Paris.
The advantage of the kriging method is that, based on the exact evaluation of individuals in
the database, it can provide an approximation of the solution as well as an error estimation for
the approximation. So, on computed points, there are obviously no errors. The points selected
for the exact evaluation will be chosen by both exploration and exploitation. Exploration
consists of searching the areas for which the error estimation of the metamodel is the largest
in order to improve it. Exploitation consists of placing yourself in the areas where the
metamodel estimates the cost function will be minimised. Figure 5 illustrates this method for
selecting individuals using a kriging metamodel. The choice will fall to points where the
minimum is most likely to be found, but also in the areas where the metamodel is the most
precise. This avoids having to converge towards a local minimum.
μ Computations
μ Computations
YES
Extraction of solutions
Parametered actions
Minimisable (objective) Constraints (obligations)
(Size variations)
Reduction of initial billet Absence of folds and no Variation on initial billet
volume underfilling dimensions
Reduction of wear on the
No constraint Variation of friction
die
Reduction of wear on the Absence of folds and no Variation on initial billet
die underfilling dimensions
Absence of folds and no
Reduction of force Variation in final press height
underfilling
Process conditions (geometry,
Reduction in damage No constraint
friction, etc.)
We note μ the number of individuals in a generation and Dim the number of parameters of an
individual.
Individual 1
Parameter 1 = 0.8
Parameter 1
(position)
Min: 0.8 Initialisation
Max:1 Individual 2
Parameter 1 = 1
Individual 1 Individual 3
Parameter 1 = 0.825 Parameter 1 = 0.958
Cost 1 = 12.099
Creation
Individual 2
Parameter 1 = 0.975 Individual 4
Cost 2 = 0.975 Parameter 1 = 0.903
Selection consists of selecting the fittest parents on which to apply the other
evolution operators. The fittest parents are those for whom the cost function
evaluated is the smallest. Crossover consists of taking the parameter sets of
selected parents and combining them to obtain the parameter sets of offspring.
Finally, mutation consists of applying, on certain offspring, modifications to
parameters based on statistic computations. This makes it possible to search a
larger range of parameter sets.
The reason for using a metamodel is that it is very simple. Without a metamodel,
an optimisation method using a genetic algorithm must evaluate a very substantial
number of individuals in order to converge towards the optimum. This number can
reach several tens of thousands. It is obvious that, in this case, this method is not
effective in terms of computation cost. A metamodel is in fact a simplified model
of the optimisation problem, that is, an approximate model of the cost function
according to optimisation parameters. This model is constructed from results
already obtained. Thus, as optimisation progresses, the metamodel becomes more
precise (Figure 9).
This figure shows that our optimisation problem has numerous extrema. They explain why
descent direction algorithms, such as BFGS, are not performance-oriented in our case: they
only find the optimum closest to their point of departure, in other words, a local optimum.
The result heavily depends on the set of parameters used for initialisation. Evolutionary
algorithms do not have this limitation, they lean towards the global minimum when the
number of individuals evaluated is sufficient.
Computations can run at the same time if the number of processors allows it.
Individual 1
Evaluation Cost 1 = 12.099
Parameter 1 = 0; 825
Individual 2
Valuation Cost 2 = 0.9750
Parameter 1 = 0.975
This means we must provide the list of individuals, that is, a table of each individual in the
experimental design. An individual is actually the list of values of each of its parameters.
These values can be either the true values between the limits of variation or those between 0
and 1 (where computation values are obtained by applying a proportionality law of 0
corresponding to the true minimum and 1 corresponding to the true maximum of the
parameter). In this case, the optimiser will evaluate the true value by using the limits that will
have been defined. This second way of providing an experimental design is useful when we
want to use an experimental design from literature or when we want to apply the same design
to several different optimisations the parameter limits of which can vary.
The experimental design must however obey user settings, namely the number of individuals
by generation and the limits of each of the parameters. Individuals on the design who have at
least one parameter with a value outside defined limits will be eliminated from the design and
will not be evaluated. If the number of individuals on the design is smaller than the number of
individuals by generation, the optimiser will determine the missing parameter sets itself. If the
number of individuals is too large, the extra individuals will not be considered.
It is possible to give the optimiser a design limited to just certain parameters. Parameters that
are not restricted by the experimental design will be automatically deduced by the optimiser.
We have defined the following individuals: (0.5,5), (10,10), (1,1), and (p1 chosen by the
optimiser,50)
In a situation like this, there is no point for the optimiser to propose a parameter set that is
impossible to obtain in reality. Discrete optimisation is used to consider these parameters
in the optimisation process. The list of values for each discrete parameter must be
provided. For the other parameters, the value will be automatically determined by the
optimiser. In practice, during optimisation, the algorithm will propose a continuous
parameter set when a new individual is created. The discrete parameter set closest to the
continuous parameter set will be retained to continue with the optimisation.
It is possible to provide the list of values for each discrete parameter. It is also possible to
provide values between 0 and 1. The true values of the parameters will then be computed
according to the limits defined by the user. This means a list of discrete values can be
reused for several types of optimisation.
An example of a list of discrete values for an optimisation on two discrete parameters and
one continuous parameter is provided below:
! example of Discrete Optimization File in 3 dimension (3
parameters)
0.5,7, ! (this is a comment)
1,7.2, ! the third parameter will be always done by the
optimizer
1.5,7.8, !
2,8.0,
which corresponds to the first parameter that takes the values 0.5, 1, 1.5, and 2, to the
second parameter that takes the values 7, 7.2, 7.8, and 8, and to a third continuous
parameter, the values of which are determined by the optimiser.
It is possible to cumulate both an experimental design and the discrete optimisation for a
same optimisation. It should also be specified that if all parameters are discrete, it is better
to define an experimental design that contains all the discrete values to launch on a single
generation since, in this case, there is no search for optimisation points.
According to a genetic algorithm as described in figure 11, the objective is to find the
following for a billet:
− the optimal diameter and length (parametered actions)
− in order to minimise its volume (the minimisable)
− with a guarantee of filling the die with no folds (the constraints).
Generation of three
Initialisation individuals
first generation Volume = 512 Good
Filled, no folds
Volume = 448
Filled, no folds
Perfect
Volume = 342
Not filled, Constraint not
no folds satisfied
Combine individuals
to obtain new "fitter" individuals
(that is, the parameters)
End after given
number of
generations
Besides results from simulations conducted, optimisation produces results specific to the
algorithm: parameter values, cost function, quantity to minimise, constraints of each
individual evaluated during the procedure
Length
Overfill (significant
Underfill flash)
Diameter
Figure 12: Example of billet dimension choices according to length and diameter
In this example, the point on the axis farthest to the left with the coordinates (4,256270;0,00)
is the best minimisable obtained (the entire die is filled and the weight is the lowest). The blue
curve represents the Pareto Frontier that corresponds to the individuals that are optimal in
their range (the fittest individuals cannot be improved without deteriorating other criteria).
C
Computation of a generation .................................................................................................................................................... 10
Constraint ................................................................................................................................................................................... 9
Cost function .............................................................................................................................................................................. 3
Creation of a population ............................................................................................................................................................. 7
D
Discrete optimisation................................................................................................................................................................ 14
E
Evolution operators .................................................................................................................................................................... 7
Evolution strategy ...................................................................................................................................................................... 4
Evolutionary algorithm............................................................................................................................................................... 4
Experimental design ................................................................................................................................................................. 13
G
Genetic algorithm ............................................................................................................................................................... 5, 7, 8
I
Initialisation of the population.............................................................................................................................................. 7, 10
K
Kriging metamodel ..................................................................................................................................................................... 5
M
Minimisable............................................................................................................................................................................ 3, 9
Mono-objective algorithm .......................................................................................................................................................... 4
O
Optimisation ............................................................................................................................................................................... 3
P
Parametered action ..................................................................................................................................................................... 9
S
Selection ................................................................................................................................................................................... 10
Selection of offspring ................................................................................................................................................................. 7
[Ejday 2011], Optimisation multi-objectifs à base de métamodèle pour les procédés de mise
en forme, Doctoral thesis, Mines Paristech, 2011