Every day in our life we can discover a lot of optimization problems. By exploring them more deeply we can reveal their fundamentals and think about their solution. The problem of the most attractive font face for the writer is such a problem. The optimization function of this problem is unknown. Beside other methods also interactive evolutionary computation could be used to solve such type of optimization problems. If we utilize interactive evolutionary computation to find solution we encounter the user fatigue problem. To eliminate this disadvantage we need to accelerate the convergence of the evolutionary computation part. In this article we propose an algorithm that tries to find a candidate solution to the most attractive font face based on input data. We assume that user likes his handwriting properties which is a subject of research for graphology. These handwriting properties will project into properties or attributes of the font face. The algorithm tries to connect Schwartz's culture model with observations and theoretical knowledge from graphology. It maps the computer font attributes to Schwartz's culture model using the revealed knowledge from graphology about handwriting. We modified our existing algorithm proposal and we included a learning phase of fuzzy membership function to match a case when user preferences are different from the theory found in graphology. The algorithm is a theoretical proposal to the user fatigue problem in interactive evolutionary computation.
We encounter many optimization problems for which we
do not know their optimization function. In such cases one
need to find such methods that can replace explicit
expression of optimization function with other resources.
Sometimes we can include human user attributes as user can
determine the optimal value according to his criteria. The
interactive evolutionary computation [
We describe the problem of the most attractive font face
for the writer which we deal with in our research in
various stages [
We have dealt with the problem of the most attractive font face in our past research and also in recent, where we discuss the question of accelerating the convergence of the interactive evolutionary computation in order to reduce the human user fatigue.
We have one type of font which face is described with a finite set of parameters. The domain of each of the parameters can be either a finite set of values or interval. The task is to set the values for these parameters. By applying values of these parameters to a given font the user should get a font face that is the most attractive for him.
To describe the font face we used the Metafont language
[
The configuration file of the font contains 62 parameters. We have experimentally chosen 25 parameters that have the largest impact on the final font face look (or 21 parameters, if we consider that the parameters are the same for upper case and lower case characters).
So we have the configuration file of the font which has a parameter vector ~p with dimension of 62. Each of the vector components is a value of one parameter and is equal to 0 if the parameter is one from the set of our chosen parameters. The modification vector4~p is also a vector of dimension 62 and his components are values of parameters that we use for the font modification. These components equal 0 that we did not chose. The new vector ~y of all parameter values we obtain if we perform vector addition of ~p and 4~p: ~y = 4~p + ~p (1) 3
According to the definition of the optimization problem
[
We have implemented and interactive evolutionary
computation methods into our application that is trying to
solve the problem of the most attractive font face [
We chose the theory of human values from Schwartz [
If we think about our values, we think what is important
for us in our life. Everyone has or can have other values
(e.g. achievement, security, benevolence) with other
degree of importance. One value can be very important for
one person but very unimportant for other. The theory of
values [
The above mentioned are features of all values. What
distinguishes one value from another is the type of goal
or motivation that the value expresses. The values theory
defines ten broad values according to the motivation that
underlies each of them. Presumably, these values
encompass the range of motivationally distinct values recognized
across cultures. According to the theory, these values are
likely to be universal because they are grounded in one
or more of three universal requirements of human
existence with which they help to cope. These requirements
are: needs of individuals as biological organisms,
requisites of coordinated social interaction, and survival and
welfare needs of groups [
Individuals cannot cope successfully with these
requirements of human existence on their own. Rather, people
must articulate appropriate goals to cope with them,
communicate with others about them, and gain cooperation in
their pursuit. Values are the socially desirable concepts
used to represent these goals mentally and the vocabulary
used to express them in social interaction. From an
evolutionary point of view these goals and the values that
express them have crucial survival significance [
Each from the ten human values 1 expresses a wide set of goals. We also know them as a culture model. Values are based on universal needs and they refer to similar concepts of values. Furthermore the values are organized by motivational similarities and oppositions. They form four main groups: • Openness to change • Self-Transcendence • Conservation • Self-Enhancement
tozChange SelfDirection
Creativity, Freedom Universalism SocialWJustice, Equality
Stimulation
ExcitingWLife Hedonism
Pleasure Achievement Success, Ambition
Power Authority, Wealth Benevolence
Helpfulness Conformity Tradition Obedience Humility
Devoutness Security
SocialWOrdre
Organizedzbyzmotivational similaritieszandzoppositions
In order to assign each human value their
characteristic font attribute values we need to search through
resources of graphology and find a connection between
human values and font attributes. We research the first level
of the basic human values. In the literature we studied
[
We consider these basic font attributes: slant (S), pressure (thickness T), spaces between characters (SPC), proportional size of middle part of font (MS), proportional size of upper part of font (BS), proportional size of lower part of font (LS), character width (SP), the shape and width of loops (TS). We try to connect these attributes with our culture model and corresponding human values: selfdirection (HSE), stimulation (HST), hedonism (HED), achievement (HAC), power (HPO), security (HBE), conformity (HCF), tradition (HTR), benevolence (HBN), universalism (HUN).
Each font in Metafont language has a different form of configuration (various equations), so does our chosen font according to table 1. In order to use this connection between human values and general font attributes we need to make another mapping between these general attributes and attributes of given font that is going to be modified. We can see in table 2 the occurrence of the font attributes by particular human values from Schwartz’s culture model. This model was tested in an European Social Survey where respondents were had to answer the questions for particular human values giving each of the values some priority (from -1 which means it is against my value to 7 which means I agree). 6
If we assume the mapping from table 2 we can propose an algorithm for estimating the font face attributes following way:
At the beginning we assume that fuzzy membership
function μ pi of attribute i which has a domain in <
mini, maxi > will be given by linear function f that
f (mini) = 0 a f (maxi) = 1. We initialize at the beginning
a set of small feed forward neural networks which learn
this linear fuzzy membership functions for every attribute
(even the simplest neural network – perceptron can learn
this 2-D linear function). Our network in comparison to
the perceptron should have three layers: input, hidden and
output layer with topology e.g. 1–5–1 in order to learn
more complex functions. We also assume we obtain the
basic human values j from the model for each user
according to his preferred weighted priority w j (e.g.: w j is
recommened by [
16 μh2 = x + 1 16 + 3. For every font attribute i we find an inverse image of an image of fuzzy membership function μ pi and we obtain value - an estimation of font attribute values of our modified font. These font attributes will determine the look of the fonts in starting sample. 4. Optionally in interactive evolutionary computation in each iteration we can introduce some form of learning for each of the fuzzy membership functions μ pi to express them also as non-linear functions. Between the iterations in the application we can teach our neural network that has learned linear functions μ pi. For the teaching we use those values of font attributes that has user evaluated as very good (they will have higher fuzzy membership function value). That way we could catch individual features of user that can lead to accelerated candidate space search of the problem of the most attractive font face. 7
We proposed a theoretical improved algorithm for font attributes estimation for our problem of the most attractive font face. We used only the first level human values from the Schwartz’s culture model, but every value can be further expanded to more sub-values. This way we could obtain more complex model of the human user which could lead to more precise estimation of font attributes so that the convergence of the evolutionary computation can be accelerated and user fatigue can be reduced. As an improvement of the algorithm we introduced the neural network that can learn individual preferences between iterations that breaks the limit of expressing each fuzzy membership function with a linear formula or any other predefined formula. The next step will be to implement the algorithm and perform experiments on real data collection. Afterwards we will be able to verify the method and determine the validity of the proposed method.
This work was partially supported by the grant – 1/0479/12 VEGA – Combinatorial structures and complexity of algorithms by the Slovak Research and Development Agency.