Method and apparatus for evolutionary design

Abstract

The invention involves generating and presenting, typically electronically, a number of design alternatives to persons who are participating in the design, selection, or market research exercise. The participants (referred to as “selectors”) transmit data indicative of their preferences among or between the presented design alternatives, and that data is used to derive a new generation of design alternatives or proposals. The new designs are generated through the use of a computer program exploiting a genetic or evolutionary computational technique. The process is repeated, typically for many iterations or cycles.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

Gene G3 represents an integer value, which makes it possible to use different crossover operators, as an alternative to the “random pick from one parent” scheme. One possibility is to compute interpolated and extrapolated values using the two values from the parents, and then to select one of these two possibilities at random. The process is described below. First, a Bernoulli trial (a “coin flip”) is performed to decide whether to interpolate of extrapolate a value for the offspring gene, from the two values of parent genes.

##STR00004##

Where γ is either a deterministic real value between 0 and 1, or a randomly generated variable within that range, for example one from a uniform distribution:
γε(0,1)
or
γ−U[0,1]
If the decision is interpolation, a formula such as the one below is used:
G3_O1=Round(μ·G3_P1+(1−μ)·G3_P2)
where μ is a real value between 0 and 1, either selected deterministically or drawn at random, at the beginning of an exercise, or at every breeding. Alternatively, different deterministic values or different distributions (in the case of variables drawn randomly) could be used at different points in the exercise. Since G3 is an integer gene, the value obtained by interpolation is rounded to the nearest integer.

If extrapolation is selected instead of interpolation, one of parent values is picked to determine the direction of such extrapolation; this is done at random. If P1's is picked, then a formula like the following one can be used:
G3_O1=Round(v·((1+μ)·G3_P1−μ·G3_P2))
where v is a (possibly random) real valued parameter, typically less than 1.0, chosen to scale down the size of the extrapolation step taken. An additional step not reflected in the formula above involves checking that the value thus computed does not exceed the allowable range for gene G3, and setting it equal to that limit if it does.

If P2 is picked as the extrapolation direction, then the following can be used:
G3_O1=Round(v·((1+μ)·G3_P2−μ·G3_P1))

The reproduction space genes, R1 and R2, being real-valued, can be treated similarly, except that the rounding operation is not needed. In the present implementation, a modified averaging operation is used, as follows:

$R{1}_{O1}=\frac{1}{2}·\left(R{1}_{P1}+R{1}_{P2}\right)+\varepsilon$
where ε is a Gaussian noise:
ε˜N(0,2)
The calculation of R2_O2 proceeds similarly.

Many other schemes are within the knowledge of those of ordinary skill in the art.

Mutation

In addition to the crossover operation, or concurrent with it, a mutation operation is applied, to introduce occasional random variation in the design candidates that are generated. In the current implementation, this is done on a gene-by-gene basis again. For each gene, a determination is made, either before of following the crossover operation, as to whether a mutation is going to be applied. This is based on Bernoulli trial with a relatively low probability of success, around 0.01 typically. In the case of categorical genes, the mutation involves selecting, at random, one of the allowable allele values, typically a value that is different from those of the two parents. In the case of integer and real-valued genes, a Gaussian noise is added to the gene value obtained after the crossover operation is complete. Again, a check is performed to make sure that the mutated value is within the allowable range; if it falls outside that range, it is set equal to the upper or lower limit, as appropriate. Another case, not used in this example (the polo shirt) is where a gene is encoded as a binary bit or string. An example would be a design feature such as a logo or rings around the sleeves, which are turned on or off, depending on whether that bit is enabled or not. In that case, a mutation would simply involve a bit flip.

Mutation, as described so far, is only applied after a breeding event, and a breeding event is only triggered by a thumbs-up vote. A refinement to the implementation is triggered when no thumbs-up votes are generated, to prevent the evolutionary process from stagnating. In that case, we generate some number of random individuals every time a voter submits a set of votes that contain no thumbs-up. The merit attributes for these random individuals are generated as described above for initializing the population. The R-space attributes for these random individuals are determined as described below, in the section that discusses re-insertion of voter “picks”.

Replacement/Removal Policies

Once one or more new design candidates (the offsprings) are created, they are introduced into the population. In order to do that, a corresponding number of current members of the population must be selected for replacement. Various strategies are employed for that purpose, ranging from purely random selection, to relatively intricate schemes based on fitness (or lack thereof) and redundancy. (Various ways used to measure redundancy and diversity are described later.) In the simple case, a population member is chosen at random: a random integer i uniformly distributed between 1 and N (the size of the population of design candidates) is generated, and the i^th member of the population is removed and replaced by the offspring. This is repeated as many times as the number of offsprings created by a mating event. Another option in the current implementation is to bias the removal by fitness, or rather, lack of fitness. In that case, a misfitness score is maintained for each member of the population, and that score is either used deterministically to remove the member(s) with the highest misfitness score(s), or stochastically by loading a “roulette wheel” with slices proportional to these misfitness scores. A very simple algorithm for computing misfitness scores, one which only relies on “thumbs-up” votes, is the following. First, any members of the population of N designs that have not been assessed yet, and that therefore have received no votes, are set aside and are not candidates for removal. This is to avoid the premature loss of design candidates, unless absolutely needed (at which point we pick uniformly at random). Next, for each of the remaining members of the population, the rate of “thumbs-ups” is computed as the ratio of “thumbs-up” votes received by that entity divided by the total number of votes received by it (i.e., the sum of “thumbs-up”, “thumbs-down”, and “neutral” votes.) Next, the average rate of “thumbs-up” for all members of the population is computed, and the population of designs is divided into two groups, those that have a “thumbs-up” rate greater then average, and those that have a rate equal to or lower than the average rate. Members of the latter group are selected at random for removal, as needed.

A more discriminating removal scheme that uses all three types of votes—thumbs-up, neutral, and thumbs-down—is sometimes used in the current implementation. In that case, the misfitness m_i for the i^th member of the population is computed as a weighted sum of that member's thumbs-up, neutral, and thumbs-down rates, as follows:
m_i=w^down·R_i^down+w^neutral·R_i^neutral+w_up·R_i^up
where the w^type terms are the weights for the particular type of vote, and R₁^type terms are the vote rates of the given type for the i^th member, with w^down>0, w^up<0, and W^neutral generally positive. For example:
m_i=3·R_i^down+1·R_i^neutral−4·R_i^up

Again, design candidates that have not been seen by any of the participants are set aside, to prevent their premature elimination (unless absolutely necessary, for example in some cases early on in an exercise.)

Another variation on the removal policy modifies the contribution to the misfitness rating of similar votes, based on whether they were all cast by the same participant or by different participants. The idea behind this version is to penalize a design candidate more if it disliked by a number of different participants, that is, if different participants gave it thumbs-down for example, as compared to when it gets the same number of thumbs-down from only one participant. In this version, the individual votes for each entity are tracked, and the misfitness is computed based on declining weighting function or schedule for each participant's votes, as in the equation that follows:

$m_i=\frac{1}{V_i}\left[w^{\mathrm{down}}·\sum _j\sum _{n=1}^{V_{i,j}^{\mathrm{down}}}{e}^{-\gamma \left(n-1\right)}+w^{\mathrm{neutral}}·\sum _j\sum _{n=1}^{V_{i,j}^{\mathrm{neutral}}}{e}^{-\gamma \left(n-1\right)}+w^{\mathrm{up}}·\sum _j\sum _{n=1}^{V_{i,j}^{\mathrm{up}}}{e}^{-\gamma \left(n-1\right)}\right]$
where m_i is the misfitness score of entity i, V_i is the total number of votes received up to that point by entity i, V_i,j^type is the number of votes of the given type cast by voter j for element

$i,\sum _j$
represents the summation over all voters j, and γ is a real parameter that determines the steepness of an exponentially decreasing weighting function that reduces the impact of additional votes cast by the same participant.

Another class of removal schemes take into account how redundant a particular member of the population is, in addition to its misfitness. The idea here is the following: given two entities that are equally unfit, it is preferable to remove the one that is genotypically similar to many other members of the population, in order to minimize the loss of genotypic diversity in the population. The redundancy computation can be based either on the reproduction genes, the feature genes, or both. These computations are described in the next section. Given a redundancy value R(P_i) for a member of the population P_i, its adjusted misfitness value m′_i is computed, as:
m_i=R(P_i)·m_i

The next section describes various ways of measuring redundancy, or its opposite, diversity.

Diversity Measurement

Diversity measurement techniques are applied to both feature genes as well as reproduction genes. We use measures of diversity to dynamically control various parameters of the evolutionary algorithm, such as the mutation rate (mutation probability), as well as various strategies used in the system, such as the removal (or replacement) strategy and the strategies used to populate a participant's voting window (which are described later.)

Redundancy

Diversity in the evolving population of N designs is measured using a metric of genotypic (or phenotypic) similarity between pairs of evolving designs (“individuals”). A pairwise similarity metric S(P_i, P_j) is defined, which returns a value between 0 and 1, where 1 signifies that P_i and P_j are genotypically (or, alternatively, phenotypically) identical. We then use this metric to compute the redundancy of each individual in the evolving population with respect to the population as a whole, as follows:

$R\left(P_i\right)=\sum _{j=1}^{N}S\left(P_i,P_j\right)$

An individual with a high redundancy value is relatively common, in the sense that there exist many other individuals in the population that are similar to it. These redundancy values are used to help maintain diversity by biasing removal policies towards more redundant individuals, as explained in more detail below. Redundancy values are also used to provide a graphical visualization of genetic (or phenotypic) diversity.

Two similarity functions are used in the current implementation. One is based on the feature genes, the other on the reproduction genes. In the case of the polo shirt, the first one uses the first three genes of the genotype. (The first two are categorical genes and the third an integer-valued gene.) We define our function S as follows:

$S\left(P_i,P_j\right)=\frac{1}{k}·\sum _k{S}^{\prime }\left(P_i^k,P_j^k\right)$
where P_i^k denotes the k^th gene of an individual i in the population.

In the case of the categorical genes, G1 and G2, S′ is given by:

$S^{\prime }\left(P_i^{1,2},P_j^{1,2}\right)=\{\begin{array}{cc}1& {\mathrm{iff}P}_i^{1,2}=P_j^{1,2}\\ 0& \mathrm{otherwise}\end{array}$

In the case of gene 3, which is an integer gene, S′ is computed as follows:

$S^{\prime }\left(P_i^3,P_j^3\right)=1-\frac{P_i^3-P_j^3}{\mathrm{Max}{\Delta }^3}$
where MaxΔ³ is the range of gene G3, that is, the difference between the maximum and minimum values it is allowed to take.

In the case of real-valued genes such as those used for the reproduction variables, redundancy or density is computed using the Euclidean distance d_ij (described earlier) in R-space between the different population members, as follows. The redundancy or density of the i^th population member is given by:

$R\left(P_i\right)=\sum _{j}f\left(d_{ij}\right)$
where d_ij is the distance in R-Space between individuals i and j, and

$f\left(x\right)=\max \left(1-\frac{x}{\mathrm{threshold}·d_{m\mathrm{ax}}},0\right)$
where threshold is a constant in the interval (0, 1] and

$d_{m\mathrm{ax}}=\underset{ij}{\max \left(d_{ij}\right)}$
Entropy

Population diversity is also measured by computing the Shannon entropy of the genotypic (or phenotypic) values in the population. A high entropy value suggests a high level of diversity. Entropy-based diversity measurement does not require a metric of similarity. We calculate the entropy of each gene independently and also combine the results using weighted averaging. To compute the entropy of a gene, we first count the frequency with which each possible allele value for that gene appears in the population. These frequencies are then plugged into the standard Shannon entropy equation:

$H\left(G_k\right)=-\sum _i\frac{M_i}{N}·{\log }_2\left(\frac{M_i}{N}\right)$
where H(G_k) is the entropy of gene G_k,

$\sum _i$
is the sum over all the different values or alleles that G_k can take, M_i is the number of occurrences in the population of the i^th allele for that gene, and N is the population size. This can be applied directly to genes G1 and G2. For genes that are similar to G1 and G2, but that span a range of many possible discrete (but ordered) values, We apply a coarse quantization to obtain a smaller set of discrete values. For genes such as gene G3 above, which span a continuous space, we convert the continuum into a set of symbols by quantizing the continuum to obtain a set of discrete bins and counting the M_i occurrences of values that fall in each of these bins.

In another possible embodiment, we may compute entropy based upon higher-order effects that occur between genes. To do this, we calculate entropy based upon the frequency with which each possible n-tuple of allele values appears in the population across the n selected genes.

Entropy, being a population-wide measure is not used when a particular member of the population is sought, as in replacement or when populating a voting window. Rather, it is used to track the evolution process, and to adjust global parameters such as the mutation probability.

Clustering

In this section, we describe the subject of clustering, which relies on similarity measurements, and which is used at different times in the embodiment described here, as discussed later. If the function S(P_i, P_j), described above, indicates the similarity between individuals P_i and P_j, then we can define a new function
D(P_i, P_j)=1−S(P_i, P_j)
to indicate the dissimilarity between these two individuals. With the function D, we can compute a dissimilarity matrix M, where each entry M_ij is the dissimilarity between individuals P_i and P_j. This matrix is symmetric and has zeros on the diagonal.

With the matrix M, we can apply any number of known clustering techniques to group the individuals either according to genotypic similarity or proximity in R-space, such as the K-medoid clustering algorithm. The K-medoid algorithm must be told the number of clusters to find. If the number of clusters that would best fit the data is not known, then the silhouette value of a clustering, can be used to decide how many clusters should be sought.

We may also cluster the human users based upon their voting behaviors. In this case, we measure the correlation in the voting records of any pair of users V_i and V_j and derive an entry M_ij in matrix M, as follows:

$M_{ij}=1-\frac{1+\mathrm{correlation}\left(V_i,V_j\right)}{2}$
Strategies for Populating the Voting Window

The voting window, also referred to as the focus window, is the window presented to each voter for the purpose of displaying a set of design candidates and collecting that voter's assessment of them. The various policies used to populate the focus window at each voting iteration are described in this section. Generally speaking, these policies seek to achieve a number of sometimes conflicting goals: a) giving the participant an opportunity to explore as much of the design space as possible, and b) giving the participant a sense that the system is responsive to his or her votes.

Voting Window Mixture Policy

The voting or focus window mixture policy examines the votes that are submitted from a first focus window and determines the number of slots in the next focus window (for the participant whose votes the system is currently processing) that will be filled with: a) offspring of design candidates shown in said first focus window, and b) samples of design candidates from the general population of design candidates.

In the present implementation, all individuals in the focus window that receive a thumbs-up vote will parent at least one, but no more than two, offspring if the number of thumbs-up votes is less than the number of focus window slots, then the individuals that have received a thumbs-up vote will be used to produce a second offspring until each has produced a second offspring, or until the slots of the new focus window are filled, whichever comes first. For example, if the focus window has six slots, and two individuals are given a thumbs-up, then both will parent two offspring, which will fill four of the six slots of the new focus window. If, instead, four individuals are given a thumbs-up, then the first two individuals will each parent two offspring, while the last two will each parent one, thus entirely filling the six slots of the focus window.

If, once all the thumbs-up votes are acted upon, any slots remain empty, then they are filled by sampling the general population of individuals, as described in the next section.

The policy described above is modified slightly when only one offspring is allowed for each candidate that receives a thumbs-up (see breeding section above.)

An alternative mixture policy used in the current implementation introduces the notion of elitism—well known in the Evolutionary Computation literature into the focus window, such that some or all of the individuals that receive a thumbs-up are retained in the next focus window. Typically, elitism is used in generational versions of evolutionary algorithms in order to avoid the disappearance of highly fit members of the population across subsequent generations. In this case, we use a similar notion in the focus window or voter window. The motivation behind that policy is to provide a sense of continuity for the participant who might be uncomfortable with the disappearance from the focus window of previously preferred design candidates. When thumbs-up voting is used, as described in this example, if more entities received thumbs-up than there are elite slots in the next window, random picks are made among those entities that received thumbs-up, until the elite slots are filled.

Yet another alternative policy in the current embodiment fixes the minimum and maximum number of focus window slots that will be allocated for: a) elites (individuals that have received a thumbs up and that are carried over), b) offspring of those individuals that have received a thumbs up, and c) samples of the general population. If the number of thumbs-up votes exceeds the number of slots allocated for offspring, then a sampling method is invoked such that only some of the recipients of thumbs-up votes are able to parent an offspring. Alternatively, we can limit the number of thumbs-up votes that a user is allowed to make per focus window. Yet another alternative is to create offspring for every individual receiving a thumbs up, but not include all the offspring in the subsequent focus window (those not appearing in the focus window will still be in the general population).

Focus Window Sampling

For focus window slots that are available for samples from the population at large, a policy is needed to decide how these candidates are chosen. In the current implementation, the simplest policy used is one where we sample randomly, uniformly across the population of individuals. This sampling takes place after all offspring (parented by the individuals that received a thumbs up) have been inserted into the population. The sampling procedure does not attempt to prevent the same individual from appearing twice in the focus window, nor does it attempt to prevent two distinct individuals that are genotypically identical from appearing together in the focus window.

An alternative approach is to bias the sampling away from regions of high redundancy (redundancy being computed as described in a previous section.) The advantage of these policies is to allow for greater exploration of the design space by the participants, by affording greater diversity in their focus windows. One such policy, used in this embodiment, utilizes R-space redundancy to discount how likely a particular population member is to be selected. More specifically, roulette wheel selection is used, with the slice given to each of the N members of the population being inversely proportional to the redundancy of that member:

$\mathrm{Pr}\left(P_i\right)=\frac{1}{N·R\left(P_i\right)}/\sum _i\frac{1}{N·R\left(P_i\right)}$

Another policy uses feature space redundancy (calculated on the basis of the feature genes) to bias the sampling, again using the same formula as above.

An alternative policy embodied in the present system performs a cluster analysis (described above) of the individuals in the population, either with respect to their positions in R-space, their genotypic characteristics, or both. Once the clusters are determined, the random sampling is conducted such that each cluster is equally likely to provide an individual for the open focus window slots, regardless of the number of individuals in each cluster. The advantage of this scheme is to allow the participant to sample equally from the different species or preference clusters (or aesthetic clusters) that are emerging during the exercise (speciation is discussed later.) This is in contrast to uniform sampling where, in effect, we sample from every cluster in proportion to the cluster size. A related approach is one where we select the representative design candidate for each cluster (the centroid or medoid of that cluster.)

In yet another policy, we bias the sampling in favor of individuals that have been infrequently viewed by that participant. In this case, the probability of a member of the population being selected is inversely related to the number of times it has appeared in his or her focus window. The probabilities used to load the roulette wheel are given by:

$\mathrm{Pr}\left(P_i\right)=\frac{1}{f\left(m_{ij}\right)}/\sum _i\frac{1}{f\left(m_{ij}\right)}$
where m_ij is the number of times that design candidate P_i has appeared in the focus window of participants, and f(x) is a monotonic function. For example:
f(m_ij)=m_ij²

In a related policy, we bias the selection in favor of individuals with feature properties that have been infrequently viewed (based on feature similarity), or in favor of individuals in regions of R-space that have been infrequently viewed in the focus window. Here too, the probabilities used to load the roulette wheel for selection are given by:

$\mathrm{Pr}\left(P_i|W^t\right)=\frac{1}{R\left(P_i|W^t\right)}/\sum _i\frac{1}{R\left(P_i|W^t\right)}$
where R(P_i|W^t), the redundancy of population member P_i with respect to the i^th focus window W^t(W¹ being the current window, W² the previous window, etc.) of the given participant is given by:

$R\left(P_i|W^t\right)=\sum _{q}S\left(P_i,W_q^t\right)$ $\mathrm{where}$ $\sum _q$
is the summation over all q members or design candidates in the focus window, and S(P_i,W_q^t) is the similarity between entity P_i and the q^th member of focus window W^t. Finally, S, the similarity function, is computed using any of the methods given in the previous section on redundancy and similarity, as appropriate.

A variation on this policy is one where we track not only the last focus window, but the last few or n focus windows and where we either give all of them equal weight or give the content of the more recent focus windows greater importance in the redundancy calculations. One particular version of this looks at the last n focus windows (n=3, e.g.), and weights them differentially. The slices or shares used in the roulette wheel in this case are given by:

$Q\left(P_i\right)=\sum _{t=1}^n{\omega }_t·\left(\frac{1}{R\left(P_i|W^t\right)}/\sum _i\frac{1}{R\left(P_i|W^t\right)}\right)$
with the weighting factors ω_t decreasing with

${\omega }_t=\frac{1}{t}$
as an example.

In yet another sampling policy, used with in this implementation, we bias the sample away from individuals that are redundant (either based on feature space similarity or on reproduction space similarity, or both) with respect to individuals that have been given a thumbs-down vote by the participant whose focus window is being populated. This is intended to minimize the chances of subjecting that participant to design candidates that he or she already voted down. This is done in a manner similar to the ones described in the previous policy, except in this case, the redundancy used is not R(P_i|W^t) but R(P_i|W^down,t), which is computed only with respect to those focus window members that received a negative vote from the participant in question. A related policy is one where we bias the sample towards individuals that are redundant (either in feature space, reproduction space, or both) with respect to individuals that have been given a thumbs-up vote (alternatively, a neutral vote) by the user whose focus window is being populated. In that case, R(P_i|W^up,t): is used, the probabilities or shares used in the roulette wheel are directly proportional to redundancy, as opposed to inversely proportional; for example:

$\mathrm{Pr}\left(P_i|W^t\right)=\frac{R\left(P_i|W^t\right)}{\sum _{i}R\left(P_i|W^t\right)}$

Yet another policy attempts to maximize the diversity in the focus window with respect to the genetic content of design candidates (either based on feature genes, reproduction genes, or both) with each subsequent sample being biased away from the properties of the individuals placed into the focus window up to that moment. The rationale is to increase diversity in the participant's focus window.

Any of the policies mentioned above, or variations thereof, can be employed to populate a participant's window when that participant returns after being away from an ongoing exercise for a while. Another policy used specifically for that purpose involves reloading a returning participant's window with the same candidates that were present in his or her last focus window when they last logged off. This policy is often problematic however, as these candidates are likely to have been removed from the population, necessitating that they be recreated and re-inserted in the population. An alternative is to present the participant with as broad a sampling of the current design population as possible. This is done by sampling from cluster representatives as described earlier. This policy is also used in the case of a participant who joins the exercise after it has been ongoing for some time, and who is not identified with any particular preference segment.

In one embodiment certain refinements are added to the voting window, which are intended to provide the participant with some or all of the following: a) a measure or indication of progress during the exercise; b) a sense of accomplishment as goal posts are reached during the exercise; c) more direct control over the evolution process; d) a sense of membership in a community of co-participants in the design process. FIG. 7C shows a voting window with two of these refinements on the right hand side. These include a progress bar 780 that covers a range from 0% to 100%, and that indicates the level of progress with a colored section. The other refinement shown in the same figure is the “pick panel” 788, which is the panel on the right hand side of the voting window, under the progress bar, labeled “Marker Designs”. In the figure, the picks panel shows three thumbnails arranged vertically, one of them with a selection in it, and the other two still blank. The picks panel displays particular design candidates at certain points during the exercise, based on one of the strategies described below. In the case shown, an “X” mark under the selected pick allows the participant to remove said pick and to restart that part of the exercise that resulted in that particular pick.

Four classes of strategies may be used in this embodiment. The first class of strategies relies on a fixed number of votes submitted by the participant; a second class depends on the degree of similarity among the candidates that are showing up in the participant's last few voting windows, and therefore may involve a variable number of voting submittals by the participant in question. A third class allows the participant to directly select one of the design candidates in the voting windows a pick, by using a special button next to the thumbs-up and thumbs-down button (not shown in this figure.) Finally, a fourth class of strategies are intended to use the pick panel to show the participant how other participants are voting.

Strategy I: Analyze a Preset Number of Votes and Pick

In this strategy, the system is set to allow each participant to view and assess a preset number n of voting windows, with typical values of n ranging between 6 and 40. In this case, the progress bar increases in proportion to the ratio of voting windows viewed by the participant up to that point, to the preset number n. After the n vote submittals, a pick is automatically made on behalf of the participant based on his voting patterns, as described below, and the progress bar is reset to zero, a new voting window populated at random from the population of designs at large, and a new set of n vote submittals is started. The voting window shown in FIG. 7C corresponds to a case where the participant is asked to go through three sets of n vote submittals, resulting in three picks.

After the preset number n of voting windows, an analysis is performed on that participant's votes on these n windows (all the votes may be examined or only the last 80% of the n submittals may be examined to remove any “training” or accommodation effects.) In one scheme, the analysis involves counting the thumbs-up votes received by each allele, and using the counts to generate the most “selected” combination of attributes values. At that point, a design candidate is assembled using these most selected attribute values, and it becomes the pick. This approach works well when there are few or no dependencies between genes. A more refined analysis that works well even if there are dependencies involves the following steps: After the n vote submittals have been received, all candidates in these voting windows that have received a positive vote (thumbs-up) are collected. Then, a first positive-vote-candidate is selected, and, starting with the first gene of that candidate, a count of how many of the other positive-vote-candidates share the same allele for that gene is performed. This is repeated for all the genes of the selected candidates, and these k counts (k being the number of genes) are added up; this count is the “representativeness” score for that candidate. This process is repeated for every one of the positive-vote-candidates, and these are ranked on the basis of their score. Of those, the top-ranking positive-vote-candidate is selected as a pick.

In one variation, the participant is given a chance to reject the chosen pick, in which case the next highest scoring one is selected as a pick, and soon. If several (for example, three) are rejected, that set of n iterations is restarted. In another variation, the participant is presented with a panel showing the three highest scoring pick candidates, and he is given the opportunity of choosing the one he deems closest to what he had been voting for.

Strategy II: Focus Window Convergence Pick

In the second class of progress indication strategies, the progress bar does not increase monotonically, but it might regress depending on the behavior of the participant. If a voter votes consistently, then it is more likely that his successive voting windows will be populated with increasingly similar design candidates; in that case, a progress bar tied to the similarity of the contents of these successive voting windows will increase. In this case, the number of vote submittals prior to a pick selection is variable. As some fraction (say, ¾) of the design candidates in the voting window became identical or very similar, the most duplicated candidate is chosen as a pick. Having made the pick, and if the pick is not rejected by the participant, a new focus window is populated (e.g., at random), and the participant starts the next phase of the process that will yield the next pick. If the pick is rejected, alternatives similar to the ones presented above under Strategy I are followed.

Strategy III: Direct Selection

In this case, after a certain number of voting submittals have been made by the participant, an additional button is enabled next to each of the design candidates in the focus window. That button is a direct pick button, which allows the participant to select the corresponding candidate to become a pick. Alternatively, when direct picks are enabled, the participant is allowed to drag the desired candidate from its location in the voting window onto the picks panel area, which will place a copy of it there. Once the participant makes a direct pick, the direct pick buttons are again disabled for a preset number of voting iterations. The pick panel has a fixed number of slots to hold the picks, and when a new pick is inserted by clicking its direct pick button, it gets placed at the top of the Pick Panel, while everything else moves down one slot, the design occupying the bottom slot being discarded. If the pick is made by dragging it onto the pick panel, then the picked design either replaces the item in the slot onto which it is dragged and dropped, or the items at that slot and below are shifted down one slot (item in bottom slot again discarded). No matter how the pick panel is managed, the history of all picks is recorded for subsequent analysis.

A variation on this scheme also allows the participant to reinsert one or more of the picks in the pick panel back into the population of design candidates (and therefore in his focus of voting window as well) later in the exercise, if the participant gets the impression that that design candidate may have been lost. In that case, the R-space values of that candidate are updated to reflect the changes that may have taken place in R-space in the interim. One cannot rely on that candidate's previous R-space coordinates to be compatible with the current configuration of R-space, since R-space is constantly in flux. A new R-space location can be chosen in one of the following ways:

• • 1) A region of R-space that contains the designs that are most similar to the design we wish to re-insert is located, based on feature gene similarity; the re-inserted design is given new R-space coordinates that place it in that neighborhood. If no designs in the population are sufficiently similar to the design to be reinserted then:
    • a) Pick R-space coordinates at random, but within the bounding box of the current population (optionally expanded by some amount);
    • b) Place the re-inserted individual in the least dense region of R- space.
    • c) Place the re-inserted individual at the periphery of the populated regions of R-space at a distance from the periphery determined by the average distance between cluster representatives.
      Strategy IV: Social Network Effect Schemes

This is a family of strategies that involve showing the participant, in a pick window, not only the pick candidates estimated based on his voting patterns, but also the picks (candidates or actual) for other voters. In this case, the most popular design candidate across voters is estimated using the same techniques described under Strategy I above, except that the positive-vote-candidates are collected from all participants, not only from the participant whose voting window we are discussing.

Speciation and Dynamic (or Co-Evolutionary) Segmentation

When the β parameter used to control mate selection (Eq. (3)) is set to a high enough value, such as 40.0, then the mechanisms and procedures outlined above will automatically allow different preference profiles to emerge and to coexist during the process. (In case Eq. (5) is used instead, then the y parameter needs to be small enough.) To the extent that the participants represent a population of consumers in a market, and to the extent that different subgroups in that market end up evolving preferences for distinct combinations of product attributes, then the system in effect performs a sort of dynamic segmentation of that market. The term “dynamic” is used here to indicate that the preference profiles and the corresponding preferred designs are co-evolved during the process. This is different from existing approaches to market segmentation, which either assume given preference profiles (for which appropriate design are developed), or given designs for which the appropriate customers are identified. This section is intended to explain how the current implementation affords that segmentation capability, and to present a simple example.

Assortative Mating

To the extent that crossover operations between certain individuals (design candidates) results in new candidates that are less preferred by the participants, we seek to prevent such mating from occurring. However, we do not know a priori which such matings will be deleterious. The R-space mechanisms that express individuals' mate choices can learn, over time, which mate pairs are compatible and which are not, based upon the assessment by participants of the outcomes of actual matings. Pairings of genetic material that are successful will gradually tend to occur more frequently and, thereby, crowd-out those pairings that are less successful. The prohibition (or reduced likelihood) of certain mate pairs is known as assortative mating, and each set of individuals that are allowed to mate with each other, but not with members of another set, is known as a species.

The evolution of species (speciation) is of direct importance to dynamic participant preference segmentation. When a design exercise begins, the R-space is homogenous: the R values of the population of design candidates are distributed uniformly in R-space. As evolution proceeds, information is gained (through the participants' feedback) about which pairings of genetic material are more successful than others. As a result of participants' assessments and the crossover operations on the reproduction genes, the distribution of the gene values in R-space becomes heterogeneous. In other words, the R-space begins to cluster. This heterogeneity is structured in a way that keeps certain individuals near each other and far from others. These clusters correspond to species, that is, sets of individuals that are reproductively isolated. As reproductive isolation emerges, each species, along with the participants who have evolved it through their voting, become specialized to a particular sub-region of the design space, and they are less subject to interference from other species.

Multiple Niches in an Ecology

When a market has multiple segments, there exists a set of distinct preference profiles for each of these segments. Each segment's preference represents an area in the design search space. These areas can be thought of as distinct ecological niches. The assortative mating dynamic allows multiple species to emerge and persist, where each species inhabits its own niche. The number of participants supporting each segment—a proxy for the size of that market segment determines the carrying capacity of that niche, and thus the size of the corresponding species. In other words, as R-space clusters form, the size of a cluster (the number of design candidates that belong to that particular species) reflects the size of the market segment (assuming a balanced level of voting among participants, which can be controlled in the current implementation, either by limiting the number of voting screens presented to each participant, or by disregarding the votes submitted by a given participant that participant has reached his or her allotted number of votes.) Because the participants discover design possibilities as they interact with the system (and thereby form opinions), and the designs evolve in response to the participants, one can describe the interaction between designs and participants to be in some sense co-evolutionary. The preferences evoked by the evolving designs allow the system as a whole to converge on a set of designs that delineate multiple segments in the market.

FIGS. 8 through 14 present an example of this dynamic segmentation process. In this example, two participants interacted with the system concurrently. The process starts with uniformly distributed reproduction genes and feature genes (see FIGS. 8 and 9, respectively) based on a random seeding of the population of candidates. After a number of voting cycles, two segments emerge, one corresponding to participant 1, and the other to participant 2. FIGS. 10 and 11 show the focus windows for the two participants at that point in the exercise. The content of each focus window is dominated by the design of choice for that participant, that is, the design choices shown to the first participant may feature different colors, patterns, and design styles (e.g. tab length) than the design choice presented to the second participant. The design choices shown to either particpant may be highly concentrated in R-space, that is, each design choice may be very similar to each other design choice shown to that particpant (e.g. similar colors, similar patterns, etc.). In other exercises the design choices presented to participants may be scattered in R-space, that is, each design choice may have a different color or pattern from other design choices being presented to the particpant. FIG. 12 shows the R-space plot at that point, with the design candidates corresponding to the two segments highlighted; in this embodiment, the two clusters are clearly distinguished. Finally, FIGS. 13 and 14 show the distribution of feature gene values for each participant at that point in the process. FIG. 13 depicts the distribution of feature genes 1 though 3 for participant 1. Style “2” is the only surviving collar style, since it is preferred by both segments. Participant 1 prefers a purplish body style (body style “1”) and a short tab length (value equal to 123).

FIG. 14 depicts a distribution of feature genes for participant 2. Collar style “2” (tab collar) is the only surviving collar style. Participant 2 prefers a green body style (body style “6”) and a long tab length (value equal to 1310).

In one embodiment, the demographic information collected about each user may be used to alter the evolutionary algorithm described above. For example, a system may accept input from a wide universe of users but only use input from a set of users having a particular demographic for the purposes of evolving the universe of design objects. This embodiment allows the manufacturer to determine the preferences of a particular market segment without requiring the manufacturer to affirmatively direct a market research effort at a particular demographic market.

In another embodiment, the system described above may be used to permit data to be gathered concerning competitive products. This is accomplished by including competitive products in the set of products designed to see if they “survive.” In one particular embodiment, the evolutionary algorithm recognizes when a competitive product is genetically similar to a set of product designs selected by one or more selectors and inserts the competitive design into the next generation of product choices.

In still another embodiment, the systems described above are used as a “virtual sales person,” that guide a consumer in determining one or more products for purchase. In this embodiment, the product designs represent the universe of items for sale by a company and successive generations of products are selected from the universe based on received user input.

In still another embodiment, the evolutionary design system includes information from commercial actors that supply raw materials to the manufacturer. For example, a supplier may provide information concerning handles available for inclusion in a product. The information typically will include dimension information and style information, but may also include pricing information. In this embodiment, a selector may be provided with information regarding the cost of a potential design and that genetic factor may be considered in creating the next generation of products for review by the selector.

In yet another embodiment, the evolutionary design techniques described above are enhanced by providing to selectors simulated endorsement data or other promotional schemes and strategies. In this embodiment, selectors that are perceived as opinion makers may have their voting preferences displayed to the voting public to determine if other selectors change their votes based on the knowledge of the opinion-makers voting preferences.

Many alterations and modifications may be made by those having ordinary skill in the art without departing from the spirit and scope of the invention. Therefore, it must be expressly understood that the illustrated embodiment has been shown only for the purposes of example and should not be taken as limiting the invention, which is defined by the following claims. The following claims are thus to be read as not only literally including what is set forth by the claims but also to include all equivalent elements for performing substantially the same function in substantially the same way to obtain substantially the same result, even though not identical in other respects to what is shown and described in the above illustrations.

Claims

1. A method of selecting a preferred one or a preferred group of forms of a product, each product form comprising a plurality of attributes, the method comprising the steps of:

(a) presenting, over an electronic network, to a plurality of selectors, one or more groups of product forms;

(b) obtaining information from a selector about the selector's preference among the presented product forms;

(c) using the obtained information to determine a derived group of product forms, each of at least some of the derived product forms comprising a combination of attributes different than the plurality of attributes of each of at least some of the presented product forms;

(d) iterating steps (a) through (c), using a derived group from step (c), until a stopping criterion is achieved; and

(e) upon achieving the stopping criterion, selecting one or a group of preferred product forms, wherein each of the attributes comprises a structural, functional, stylistic, or economic feature of the product.

2. The method of claim 1 wherein the stopping criterion is achieved when the derived group of product forms ceases to change significantly after an iteration.

3. The method of claim 1 wherein step (b) comprises obtaining information indicative of which presented product forms are preferred by the selector.

4. The method of claim 1 wherein step (b) comprises obtaining information indicative of which presented product forms are not preferred by the selector.

5. The method of claim 1 wherein step (b) comprises obtaining information indicative of relative preference of the selector from among the presented product forms.

6. The method of claim 1 wherein the obtained information from step (b) includes information indicative of the confidence of the selector in the selector's preference.

7. The method of claim 1 wherein step (b) comprises obtaining information indicative of a rating assigned to at least some of the presented product forms by at least one selector.

8. The method of claim 1 wherein step (b) comprises obtaining information indicative of a preference as between a presented product form and a previously presented product form.

9. The method of claim 1 wherein different selectors are presented with different groups of product forms.

10. The method of claim 1 wherein each selector comprises a person or a group of persons.

11. The method of claim 1 wherein step (c) includes determining the derived group of product forms by selecting the derived group of product forms.

12. The method of claim 1 wherein step (c) includes determining the derived group of product forms by selecting the derived group of product forms.

13. The method of claim 1 wherein step (c) includes determining the derived group of product forms by generating the derived group of product forms using a computational algorithm.

14. The method of claim 13 wherein the computational algorithm comprises an evolutionary algorithm.

15. The method of claim 13 wherein the computational algorithm comprises a genetic algorithm.

16. The method of claim 1 wherein, for each iteration, each selector is presented with a group of product forms substantially different from the groups presented to the other selectors.

17. An electronic network comprising computers for use by selectors to express preferences for certain forms of a product, each product form comprising a plurality of attributes, the network being configured to:

(a) present one or more groups of product forms to a plurality of selectors;

(b) obtain data from a selector indicative of the selector's preference from among the presented product forms;

(c) use the obtained data to determine a derived group of product forms, each of at least some of the derived product forms comprising a combination of attributes different than the plurality of attributes of each of at least some of the presented product forms;

(d) iterate steps (a) through (c), using a derived group from step (c), until a stopping criterion is achieved; and

(e) upon achieving the stopping criterion, output data informing a decision to select one or a group of preferred product forms, wherein each of the attributes comprises a structural, functional, stylistic, or economic feature of the product.

18. The network of claim 17 wherein the stopping criterion comprises convergence of a group of product forms.

19. The network of claim 17 further configured to identify a subset of the plurality of selectors having similarity among expressed product form preferences.

20. The network of claim 17 configured to obtain data indicative of relative preference of the selector from among presented product forms.

21. The network of claim 17 configured to obtain data indicative of the confidence in an expressed selector preference.

22. The network of claim 17 configured to obtain data indicative of a rating assigned to at least some of the presented product forms.

23. The network of claim 17 configured to obtain data from a selector indicative of the selector's preference as between a presented product form and a previously presented product form.

24. The network of claim 17 configured to present different selectors with a different group of product forms.

25. The network of claim 17 configured to use a computational algorithm to generate the derived group of product forms.

26. The network of claim 17 configured to use selection to assemble the derived group of product forms.

27. The network of claim 17 configured to use an evolutionary algorithm to generate the derived group of product forms.

28. The network of claim 17 configured to use a genetic algorithm to generate the derived group of product forms.

29. The network of claim 17 wherein, for each iteration, any one selector is presented with a group of product forms substantially different from the groups presented to the other selectors.

30. The network of claim 17 further comprising at least one server for performing at least function (c).

31. A method of selecting one or more preferred forms of a product, the method comprising the steps of:

(a) presenting, over an electronic network, to a plurality of selectors, one or more groups of the product forms;

(b) obtaining information from a selector about the selector's preference among the presented product forms;

(c) using the obtained information to determine one or more derived product forms;

(d) repeating steps (a) through (c), using at least some of the one or more derived product forms from step (c), until a stopping condition is met; and

(e) upon achieving the stopping condition, selecting one or more of the remaining derived product forms as the one or more preferred product forms, wherein the product comprises a mass produced good, a consumer good, a manufactured good, a service, advertising material, or packaging material.

32. The method of claim 31 wherein the product comprises apparel, footwear, a computer, a telephone, a chair, a seat, an automobile, a bicycle, a home, a building, a boat hull, or a billboard.

33. The method of claim 31 wherein step (b) comprises obtaining information indicative of which product forms among the presented product forms are preferred by the selector.

34. The method of claim 31 wherein step (b) comprises obtaining information indicative of which product forms among the presented product forms are not preferred by the selector.

35. The method of claim 31 wherein different selectors are presented with different groups of product forms.

36. The method of claim 31 wherein each selector comprises a person or a group of persons.

37. The method of claim 31 comprising using a genetic or evolutionary computational technique to determine the derived product forms.

38. A system for selecting one or more preferred forms of a product, the method comprising the steps of:

(a) means for presenting, over an electronic network, to a plurality of selectors, one or more groups of the product forms;

(b) means for obtaining information from a selector about the selector's preference among the presented product forms;

(c) means for using the obtained information to determine one or more derived product forms;

(d) means for repeating the functions of means (a) through (c), using at least some of the one or more derived product forms from step (c), until a stopping condition is met; and

(e) means for, upon achieving the stopping condition, selecting one or more of the remaining derived product forms as the one or more preferred product forms, wherein the product comprises a mass produced good, a consumer good, a manufactured good, a service, advertising material, or packaging material.

39. The system of claim 38 wherein the product comprises apparel, footwear, a computer, a telephone, a chair, a seat, an automobile, a bicycle, a home, a building, a boat hull, or a billboard.

40. A system for selecting one or more preferred forms of a product, the system comprising one or more server computers for communicating with a plurality of client computing devices over a computer network, each of the client computing devices being adapted for use by one of a plurality of selectors, the system being configured: to present at least one group of product forms to each of the selectors; to obtain information about the selector's preferences among the presented product forms; to use the obtained information to determine one or more derived product forms; and to repeat the presenting, obtaining, and using functions with at least one of the one or more derived product forms until a stopping condition is met, and then to select one or more of the remaining derived product forms as the one or more preferred product forms, wherein the product comprises a mass produced good, a consumer good, a manufactured good, a service, advertising material, or packaging material.

41. The system of claim 40 wherein the products comprises apparel, footwear, a computer, a telephone, a chair, a seat, an automobile, a bicycle, a home, a building, a boat hull, or a billboard.

42. A method of identifying one or more preferred product forms, the method comprising the steps of:

(a) presenting, over an electronic network, to a plurality of selectors, one or more groups of the product forms;

(b) obtaining information from a selector about the selector's preference among the presented product forms;

(c) using the obtained information to determine at least one derived group of the product forms;

(d) repeating steps (a) through (c), using the at least one derived group from step (c), until a stopping condition is met; and

(e) after achieving the stopping condition, identifying one or more of the remaining derived product forms as the one or more preferred product forms, manufacturing at least one of the one or more preferred product forms or a derivative thereof, and effecting a sale of the at least one of the one or more preferred product forms or the derivative thereof.

43. The method of claim 42 wherein step (b) comprises obtaining information indicative of which product forms among the presented product forms are preferred by the selector.

44. The method of claim 42 wherein step (b) comprises obtaining information indicative of which product forms among the presented product forms are not preferred by the selector.

45. The method of claim 42 wherein different selectors are presented with different groups of product forms.

46. The method of claim 42 wherein each selector comprises a person or a group of persons.

47. The method of claim 42 comprising using a computational technique to determine the at least one derived group of the product forms.

48. A system for identifying one or more preferred product forms, the method comprising the steps of:

(a) means for presenting, over an electronic network, to a plurality of selectors, one or more groups of the product forms;

(b) means for obtaining information from a selector about the selector's preference among the presented product forms;

(c) means for using the obtained information to generate or select at least one derived group of the product forms;

(d) means for repeating the functions of means (a) through (c), using the at least one derived group from step (c), until a stopping condition is met; and

(e) means for, after achieving the stopping condition, identifying one or more of the remaining derived product forms as the one or more preferred product forms and effecting a sale of at least one of the one or more preferred product forms.

49. A method of identifying a decision object from a plurality of product forms, each product form having a market value, the method comprising the steps of:

(a) presenting, over an electronic network, to a plurality of selectors, one or more groups of the product forms;

(b) obtaining information from a selector about the selector's preference among the presented product forms;

(c) using the obtained information to determine a derived group of product forms; and

(d) repeating steps (a) through (c), using a derived group from step (c), to arrive at a product form with an optimized market value.

50. A method of selecting one or more preferred forms of a product or service, each form of product or service comprising a plurality of attributes, the method comprising the steps of:

(a) presenting, over an electronic network, to a plurality of selectors, one or more groups of the forms of the product or service;

(b) obtaining information from a selector about the selector's preference among the presented forms of the product or service;

(c) using the obtained information to determine one or more derived forms of the product or service;

(d) iterating steps (a) through (c), using derived forms from step (c), until a stopping criterion is achieved; and

(e) upon achieving the stopping criterion, selecting one or a group of preferred forms of the product or service for further development, manufacture, use, or sale.

51. The method of claim 50 wherein each of the attributes comprises a structural, functional, stylistic, or economic feature of the product or service.

52. The method of claim 50 wherein each of the attributes comprises an element of the product or service that is optional.

53. The method of claim 50 wherein each of the attributes comprises an element of the product or service that has many values or many subtypes.

54. A plurality of client computers for use in selecting one or more preferred forms of a product or service, each form of product or service comprising a plurality of attributes, each of at least some of the client computers configured to:

(a) receive for presentation, to one of a plurality of selectors, one or more groups of the forms of the product or service;

(b) obtain information from the selector about the selector's preference among the presented forms of the product or service, and transmit at least some of that information to at least one server;

(c) receive for presentation to the selector one or more derived forms of the product or service, the one or more derived forms having been determined by use of the transmitted information;

(d) iterate functions (a) through (c), using derived forms from (c), until a stopping criterion is achieved by the at least one server; and

(e) upon the at least one server achieving the stopping criterion, receive for presentation to the selector one or a group of preferred forms of the product or service for manufacture, use, further development, or sale.

55. The plurality of client computers of claim 54 wherein each of the attributes comprises a structural, functional, stylistic, or economic feature of the product or service.

56. The plurality of client computers of claim 54 wherein each of the attributes comprises an element of the product or service that is optional.

57. The plurality of client computers of claim 54 wherein each of the attributes comprises an element of the product or service that has many values or many subtypes.

58. A method conducted by a designer for selecting a preferred one or a preferred group of products or services from a larger population of possible products or services, each of the products or services including a plurality of combinations of attributes, the method comprising causing a third party to:

(a) present, over an electronic network, to a plurality of selectors, one or more groups of the products or services;

(b) obtain information from a selector expressing a selector preference among the presented products or services;

(c) use the information to obtain a derived group of products or services; and

(d) iterate actions (a) through (c), using a derived group from step (c), and, upon achieving a stopping criterion, select one or a group of preferred products or services for submission to the designer for further development, manufacture, use, or sale.

59. A method to identify a preferred product form from a plurality of product forms, the method comprising:

(a) presenting, with a computing device over an electronic network, ones of the plurality of product forms via voting window presentation slots;

(b) designating a first portion of the voting window presentation slots for offspring products;

(c) designating a second portion of the voting window presentation slots for unviewed ones of the plurality of product forms;

(d) obtaining, with the computing device, respondent preference information from the voting window presentation slots associated with respective ones of the plurality of product forms;

(e) when the obtained preference information for a respective voting window is associated with a positive score, reducing a number of computation iterations by generating, by applying generative grammar techniques with an evolutionary algorithm with the computing device, an offspring product form to populate a respective one of the first portion of the voting window slots during a subsequent presentation;

(f) when the obtained preference information for the respective voting window is associated with a negative score, replacing, with the computing device, a respective one of the second portion of the voting window slots with an alternate one of the plurality of product forms not previously viewed for presentation during the subsequent presentation;

(g) iterating, with the computing device, (d) through (f) until a stopping condition occurs; and

(h) identifying, with the computing device, the preferred product form having a greatest preference information value.

60. A method as defined in claim 59, wherein designating the second portion of the voting window presentation slots for offspring products increases a voting diversity of the respondent.

61. A method as defined in claim 59, wherein the positive score is identified by a thumbs-up vote and the negative score is identified by a thumbs-down vote.

62. A method as defined in claim 59, wherein generating the offspring product form is based on two or more products having the positive score.

63. A method as defined in claim 59, wherein the alternate one of the plurality of product forms has not previously been presented via the voting window presentation slots.

64. A method as defined in claim 59, further including retaining the product form associated with the positive score along with the offspring product form when presenting a subsequent iteration of the number of voting window presentation slots.

65. A method as defined in claim 59, further including increasing a respondent product viewing diversity by restricting a number of positive score votes during each presentation iteration of the voting window presentation slots.

66. A method as defined in claim 59, further including decreasing a number of iterations to identify the indication of convergence by applying generative grammar techniques with the evolutionary algorithm.

67. A method as defined in claim 59, wherein the stopping condition includes an indication of convergence of generated offspring product forms.

68. A system comprising a server computer communicatively connected to one or more client computing devices over a network to identify a preferred product form from a plurality of product forms, the system to, at least:

(a) present ones of the plurality of product forms via voting window presentation slots;

(b) designate a first portion of the voting window presentation slots for offspring products;

(c) designate a second portion of the voting window presentation slots for unviewed ones of the plurality of product forms;

(d) obtain respondent preference information from the voting window presentation slots associated with respective ones of the plurality of product forms;

(e) when the obtained preference information for a respective voting window is associated with a positive score, reduce a number of computation iterations by generating, by applying generative grammar techniques with an evolutionary algorithm, an offspring product form to populate a respective one of the first portion of the voting window slots during a subsequent presentation;

(f) when the obtained preference information for the respective voting window is associated with a negative score, replace a respective one of the second portion of the voting window slots with an alternate one of the plurality of product forms not previously viewed for presentation during the subsequent presentation;

(g) iterate (d) through (f) until a stopping condition occurs; and

(h) identify the preferred product form having a greatest preference information value.

69. A system as defined in claim 68, wherein designating the second portion of the voting window presentation slots for offspring products increases a voting diversity of the respondent.

70. A system as defined in claim 68, wherein the positive score is identified by a thumbs-up vote and the negative score is identified by a thumbs-down vote.

71. A system as defined in claim 68, wherein generating the offspring product form is based on two or more products having the positive score.

72. A system as defined in claim 68, wherein the alternate one of the plurality of product forms has not previously been presented via the voting window presentation slots.

73. A system as defined in claim 68, wherein the system is to retain the product form associated with the positive score along with the offspring product form when presenting a subsequent iteration of the number of voting window presentation slots.

74. A system as defined in claim 68, wherein the system is to increase a respondent product viewing diversity by restricting a number of positive score votes during each presentation iteration of the voting window presentation slots.

75. A system as defined in claim 68, wherein the system is to identify the stopping condition via an indication of convergence of generated offspring product forms.

76. A method to decrease a number of computation iterations when identifying a preferred product form from a plurality of product forms, the method comprising:

(a) presenting, with a computing device over an electronic network, ones of the plurality of product forms via voting window presentation slots;

(b) designating a first portion of the voting window presentation slots for offspring products;

(c) designating a second portion of the voting window presentation slots for unviewed ones of the plurality of product forms;

(d) obtaining, with the computing device, respondent preference information from the voting window presentation slots associated with respective ones of the plurality of product forms;

(e) when the obtained preference information for a respective voting window is associated with a positive score, reducing the number of computation iterations by generating, by applying generative grammar techniques with an evolutionary algorithm via the computing device, an offspring product form to populate a respective one of the first portion of the voting window slots during a subsequent presentation;

(f) when the obtained preference information for the respective voting window is associated with a negative score, replacing, with the computing device, a respective one of the second portion of the voting window slots with an alternate one of the plurality of product forms not previously viewed for presentation during the subsequent presentation;

(g) iterating, with the computing device, (d) through (f) until a stopping condition occurs; and

(h) identifying, with the computing device, the preferred product form having a greatest preference information value.

77. A method as defined in claim 76, wherein designating the second portion of the voting window presentation slots for offspring products increases a voting diversity of the respondent.

78. A method as defined in claim 76, wherein the positive score is identified by a thumbs-up vote and the negative score is identified by a thumbs-down vote.

79. A method as defined in claim 76, wherein generating the offspring product form is based on two or more products having the positive score.

80. A method as defined in claim 76, wherein the alternate one of the plurality of product forms has not previously been presented via the voting window presentation slots.

81. A method as defined in claim 76, further including retaining the product form associated with the positive score along with the offspring product form when presenting a subsequent iteration of the number of voting window presentation slots.

82. A method as defined in claim 76, further including increasing a respondent product viewing diversity by restricting a number of positive score votes during each presentation iteration of the voting window presentation slots.