TECHNICAL RESEARCH REPORT

Using Neural Networks to Generate Design Similarity Measures

by Sundar Balasubramanian, Jeffrey W. Herrmann

T.R. 99-38

I R
INSTITUTE FOR SYSTEMS RESEARCH

ISR develops, applies and teaches advanced methodologies of design and analysis to solve complex, hierarchical,
heterogeneous and dynamic problems of engineering technology and systems for industry and government.
ISR is a permanent institute of the University of Maryland, within the Glenn L. Martin Institute of Technol-
ogy/A. James Clark School of Engineering. It is a National Science Foundation Engineering Research Center.
Web site [Link]
 Using Neural Networks to Generate Design Similarity Measures
for Variant Fixture Planning
Sundar Balasubramanian and Jeffrey W. Herrmann

Department of Mechanical Engineering and Institute for Systems Research

University of Maryland
College Park, Maryland 20742

June 17, 1999

Abstract
This paper describes a neural network-based design similarity measure for a variant
fixture planning approach. The goal is to retrieve, for a new product design, a useful
fixture from a given set of existing designs and their fixtures. However, since calculating
each fixture’s feasibility and then determining the necessary modifications for infeasible
fixtures would require too much effort, the approach searches quickly for the most
promising fixtures. The proposed approach uses a design similarity measure to find
existing designs that are likely to have useful fixtures. The use of neural networks to
generate design similarity measures is explored. This paper describes the back-
propagation algorithm for network learning and highlights some of the implementation
details involved. The neural network-based design similarity measure is compared
against other measures that are based on a single design attribute.
Keywords: Neural networks, variant fixture planning, design similarity.

1. Problem Statement
This paper describes a neural network-based design similarity measure for a variant
fixture planning approach. The goal is to retrieve, for a new product design, a useful
fixture from a given set of existing designs and their fixtures. However, since calculating
each fixture’s feasibility and then determining the necessary modifications for infeasible
fixtures would require too much effort, the approach searches quickly for the most
promising fixtures. The objective is to determine a design similarity measure that reflects
fixture-usefulness. In mathematical terms, given a new design D, the neural network
model is required to estimate usefulness of an existing fixture (of design D′) for design D.
Input to the neural network model are the design attributes for both designs D and D′,
while the output is given as follows:
Q′
Output (Relative Metric) = (1.1)
Q
where Q′ is the usefulness metric for new design D with existing fixture (of design D′ )
and Q is the usefulness metric for new design D with its best generatively designed
fixture.

The best generatively designed fixtures provide the least maximum contact reaction
force for an applied unit torque (clockwise or counter-clockwise). The reciprocal of this
reaction force is called the torque resistance metric [Bro96], which is the measure of
quality and usefulness of a fixture for a given design, in our discussion. If the maximum
contact reaction force is smaller, the metric is larger. Therefore, for any design, the best
generatively designed fixture yields the highest quality metric. This implies that the
output for the neural network model lies in the range [0, 1].

In planar fixture synthesis, a design can be characterized by design attributes such as

MID, Area and MVED. The Maximum Inter-vertex Distance (MID) for a part (or design)
is the maximum length between any pair of vertices for the polygon that represents the
2D projection of a design. Similarly, the Area measure is the area covered by the polygon
representing the design and the Maximum Vertex-Edge Distance (MVED) is the
maximum perpendicular distance between any vertex and an edge for the polygon that
represents the 2D projection of a design.

A design similarity measure based on any single design attribute like MID cannot
represent fixture-usefulness accurately. However, a measure that uses a collection of
these design attributes might together provide a better measure of fixture-usefulness.
Note that parameters such as number of sides or number of edges do not reflect fixture-
usefulness; for example – an arc in the 2D projection of a part can be approximated by
chords, in which case the number of edges loses its significance as a design attribute.
This paper describes a method that establishes a mapping between the design attributes
mentioned above and fixture usefulness.

2. Background and Approach

2.1 Variant Fixture Planning

A variant fixture planning approach exploits existing knowledge by retrieving, for a new
product design, a useful fixture from a given set (or database) of existing designs and
their fixtures. It will enable a manufacturing firm to reuse dedicated fixtures by
identifying an existing fixture that requires only a minor change to become an effective
fixture for a new design. This will reduce the amount of time spent constructing fixtures.
However, since calculating each fixture’s feasibility and then determining the necessary
modifications for infeasible fixtures would require too much effort, the approach searches
quickly for the most promising fixtures.

As explained below, the proposed approach uses a design similarity measure to find
existing designs that are likely to have useful fixtures. The design similarity measure
allows the approach to identify the most promising fixtures (those that correspond to the
most similar designs) and process only those in more detail.

2
 A particular class of products and modular components has been considered. One
face of the part rests on the supporting plane (a baseplate) and the fixture elements
constrain all motion of the part in the supporting plane. Thus, only the 2D projection of
any given design onto the supporting plane is considered for fixture planning. Only
polygonal shapes are considered. In this setting, a fixture is a set of three locators (pins)
and one clamp. Generative fixture planning approaches [Bro96, Zhu96] have been
developed for this domain. However, because they create a unique fixture for each
design, generative approaches do not support the need to reuse fixtures.

Given a new design D, and an existing design D′, we define the fixture-based design
similarity measure h(D’,D)
 Attr ( D ) 
− Abs  − 1 
 Attr ( D ′ ) 
=e if Attr(D) ≥ 0.9*Attr(D’)

=0 otherwise

where Attr() represents any one of the design attributes MID, Area, and MVED.

Note that this design similarity measure follows the approach described in
Balasubramanian et al. [Bal98] and is based on ideas from Herrmann and Singh [Her97].
Specifically, the design similarity measure reflects the underlying fixture usefulness
because these attributes are related to fixture usefulness. In addition, the measure is not
symmetric. However, these measures based on any single design attribute do not yield a
consistent measure, as shown in Section 3. The remainder of this paper describes a neural
network-based design similarity measure that utilizes these design attributes and
establishes a mapping between the design attributes and fixture usefulness.

2.2 Introduction to Neural Networks

Artificial neural networks (ANNs) are computer simulations of biological neurons,
composed of nonlinear computational elements operating in parallel. Neural networks
consist of nodes (neurons) and synaptic connections that connect these nodes. Each
connection is assigned a relative weight, also called connection strength, or synaptic
strength. The output at each node depends on the threshold (bias or offset) specified and a
transfer (activation) function.

In mathematical terms, a neural network model can be represented by the following

constituent parameters [Mul90]:

1. A state variable si is associated with each node i.

2. A weight wij associated with a connection between a node i and, a node j that the node
i is connected to, such that signals flow from j to i.
3. A bias vi associated with each node i.
4. A transfer function fi(sj, wij, vi) for each node i, which determines the state of the node
as a function of the summed output from all the nodes that are connected to node i,
and its bias.

3
The state variable si of a node i is given by,

 J 
si = f i  ∑ wij s j − vi  (1.2)
 j =1 

Note that transfer functions and bias terms are absent for the input nodes. The
sigmoid function is the commonly used transfer function. Another popular activation
function is the ‘tanh’ function. These functions are monotonic with a finite derivative.
The sigmoid function is equivalent to the ‘tanh’ activation function if we apply a linear
~
transformation α~ = α / 2 to the input and a linear transformation f = 2 f − 1 to the output.
The sigmoid and ‘tanh’ functions are given as follows:
1
Sigmoid function = f (α ) = (1.3)
(1 + e − 2 βα )
e βα − e − βα
‘tanh’ function = f (α ) = (1.4)
e βα + e − βα

where β determines the steepness. Generally, β is set to unity, which results in a sigmoid
transfer function of the form,

1
f (α ) = (1.5)
(1 + e −α )

If yi = f (hi ), where hi is the summed input to a node i, f ′(hi ) = y i (1 − y i ) . In other

words, the sigmoid function lends itself to back-propagation since corrections to weights
can be obtained in terms of the state values of the output nodes. This is discussed in detail
in Section 2.3.

Some typical applications of neural networks involve pattern mapping, pattern

completion and pattern classification. Interest in neural networks stem from the fact that
these models are capable of performing complex tasks that are impossible with sequential
models. Neural networks are particularly useful in problems where the logical structure
or the input-output relation is poorly understood. Neural networks are capable of learning
and generalizing from examples in the absence of explicit rules or an analytical structure.

Classification of Neural Network Models

A neural network is initialized with a random set of weights. Adjustment of the
connection weights to improve a predefined performance measure of a neural network is
called learning. There are two types of learning methods: supervised and unsupervised.
In the case of supervised learning, a set of training input vectors and their associated
output vectors are presented to adjust the weights in the network. In unsupervised
learning, no output vectors are specified. Strategies such as those based on penalty and
reward, and genetic algorithms are employed to adjust the synaptic strengths.

4
 Neural nets can also be classified into feed-forward and feed-back networks, on the
basis of direction in which signals flow. In a feed-forward network, signals flow in only
one direction, from the input layer through the intermediate hidden layers to the output
layer. Neurons in the same layer do not communicate with each other. In a feedback
network, signals can flow from the output of any node to the input of any node.

Neural nets can have binary inputs or continuous valued inputs. A neural network
model that consists of an input neuron layer that feeds directly into an output layer is
termed as a simple perceptron. However, neural network models can possess inner, or
hidden, neuron layers that intervene between the input and output layers. Such models are
called multi-layered networks. The best-studied class of layered neural networks is the
feed-forward network. Within layered networks, there are fully-connected and partially
connected networks. In a fully-connected network, all the nodes in each layer are
connected to all the nodes in the previous layer, while this criterion does not hold for a
partially connected neural network topology. See [Lip87] for a neural net taxonomy.

Multi-layered Networks
Multi-layered are networks with input, output, and inner (or hidden) neuron layers that
intervene between the input and output layers. The input-output relation defines a
mapping, and the neural network provides a representation of this mapping. The number
of hidden layers and number of nodes in each hidden layer depend on the complexity of
the problem, and to a large extent vary from problem to problem. Increasing the number
of hidden layers increases the complexity of a neural network, and may or may not
enhance the performance of the network. In most cases, more nodes in a hidden layer
result in a better network performance but lead to a longer training time. Based on
previous experience, one or two hidden layers provide a better performance [Hua99],
while not requiring extensive training time. Sensitivity analysis can be performed to
obtain an optimal model structure, by varying the number of hidden layers and the
number of nodes per layer and evaluating performance for each of these alternative
models.

Three-layered Feed-forward Network

A three-layer feed-forward network consists of an input layer, an output layer and a
hidden layer, resulting in two layers of weights – one connecting the input to the hidden
layer and the other connecting the hidden layer to the output layer. Figure 1 illustrates a
three-layered feed-forward implementation architecture for evaluating design similarity
measures with MID, Area and MVED parameters for both the existing and the new
design as inputs and the relative metric as the output. (Note that not all weights are
shown).

The states of nodes in the various layers are as follows:

K
Hidden layer: y j = f (h j ) h j = ∑ w jk x k for j = 1, …, J (1.6)
k =1
J
Output layer: yi = f (hi ) hi = ∑ wij y j for i = 1, …, I (1.7)
j =1

5
 Relative Metric i – Output Layer

wij

j – Hidden Layer
Bias

w jk

k – Input Layer

Bias MIDExist AreaExist MVEDExist MIDNew AreaNew MVEDNew

Figure 1. A three-layered feed-forward network

Note that the subscript i refers to one of the I output nodes, the subscript j refers to
one of the J nodes in the hidden layer (including the bias node), and the subscript k refers
to one of the K input nodes (including the bias node).

The topology presented in Figure 1 is equivalent to the description in Section 2.2. The
bias vi associated with each node i is eliminated and instead, an additional bias node is
added to the input and hidden layers. From Equation (1.2),

 J   J   J +1 
si = f i  ∑ wij s j − vi  = f i  ∑ wij s j − wi ( j +1)  = f i  ∑ wij s j  (1.8)
 j =1   j =1   j =1 

where wi(j+1) is the weight associated with the connection between the bias node j+1 (that
feeds into node i) and node i. The bias nodes always have a value of 1. In other words,
σ 0 = y 0 = 1 . The bias nodes act in a manner equivalent to the intercept term in regression
models. In the example shown in Figure 1, there are seven nodes (which includes the bias
node) in the input and hidden layers. The output layer has only one node.

2.3 The Back-Propagation Algorithm

Learning is accomplished through an adaptive procedure, known as a learning rule or
algorithm. Learning algorithms indicate how weights should be incrementally adapted to
improve a predefined performance measure. Learning can be viewed as a search in a
multidimensional weight space for a solution, which gradually optimizes a predefined
objective function [Has95]. Back-propagation is the most extensively used training
method. The back-propagation algorithm is an iterative gradient method based algorithm
developed to introduce synaptic corrections (weight adjustments) by minimizing the sum
of squared error (objective function).

6
Back-Propagation in a Three-layered Feed-forward Network
The aim is to choose weights such that the output deviation function is minimized. The
output deviation function (sum of squared error) is given as follows:

1 I
D= ∑ [d i − yi ]2 (1.9)
2 i =1

where i represents an output node, yi = Actual network output; di = Desired output.

Adopting the method of steepest descent, we can determine corrections to weights as

follows [Mul90] (considering connections for the output nodes):

wij (t + 1) = wij (t ) + δwij (1.10)

∂D ∂h
δwij = −ε = ε [d i − yi ]f ′(hi ) i (1.11)
∂wij ∂wij
where ε = learning rate.

For the sigmoidal transfer function [given yi = f (hi ) ],

f ′(hi ) = yi (1 − yi ) (1.12)
∂hi
= yj (1.13)
∂wij

Substituting (1.12) and (1.13) in (1.11), we get,

δwij = ε [d i − yi ]yi (1 − yi ) y j = ε ∆ i y j (1.14)

where ∆ i = [d i − yi ]yi (1 − yi )

For weights associated with synaptic connections between the input and hidden layer,

w jk (t + 1) = w jk (t ) + δw jk (1.15)
∂D I
∂h ∂y j
δw jk = −ε = ε ∑ [di − yi ]f ′(hi ) i (1.16)
∂w jk i =1 ∂y j ∂w jk
∂D I ∂h j
δw jk = −ε = ε ∑ [d i − y i ]f ′(hi ) wij f ′(h j ) (1.17)
∂w jk i =1 ∂w jk

For a sigmoid transfer function,

f ′(h j ) = y j (1 − y j ) (1.18)

7
 ∂h j
= xk (1.19)
∂w jk

Substituting (1.18) and (1.19) in (1.17), we get:

I
δw jk = ε ∑ [d i − y i ] y i (1 − y i ) wij y j (1 − y j ) x k (1.20)
i =1
I
δw jk = εy j (1 − y j ) x k ∑ [d i − y i ]y i (1 − y i ) wij (1.21)
i =1
I
δw jk = εx k ∆ j where ∆ j = y j (1 − y j )∑ ∆ i wij (1.22)
i =1

Convergence can be accelerated by adding a momentum term as follows [Lip87]:

wij (t + 1) = wij (t ) + δwij + η ( wij (t ) − wij (t − 1)) (1.23)

w jk (t + 1) = w jk (t ) + δw jk + η ( w jk (t ) − w jk (t − 1)) (1.24)

where η = momentum factor and 0 < η < 1.

In the region of weight space for which the error surface has relatively low curvature,
it can be shown [Bis95] that the momentum term aids in increasing the learning rate ε to
an effective learning rate of ε /(1-η).

This recursive method of back-propagation can be further extended for networks with
more than one hidden layer. The values for the parameter constants such as ε and η can
be altered to either improve performance or to reduce computational effort. This is further
discussed in Section 2.6. There is no general criterion to determine these constants. The
optimal values depend on the problem at hand.

Back-Propagation in a Four-layered Feed-forward Network

Here, the subscript j refers to one of the J nodes in the second hidden layer, subscript k
refers to one of the K nodes in the first hidden layer, and the subscript l refers to one of
the L input nodes.

L
1st Hidden layer: yˆ k = f (hˆk ) hˆk = ∑ wˆ kl xl for k = 1, …, K (1.25)
l =1
K
2nd Hidden layer: y j = f (h j ) h j = ∑ w jk yˆ k for j = 1, …, J (1.26)
k =1
J
Output layer: yi = f (hi ) hi = ∑ wij y j for i = 1, …, I (1.27)
j =1

8
 For weights associated with synaptic connections between the input and first hidden
layer,

wˆ kl (t + 1) = wˆ kl (t ) + δwˆ kl (1.28)
J
δwˆ kl = ε∆ˆ k xl where ∆ˆ k = yˆ k (1 − yˆ k )∑ ∆ j w jk (1.29)
j =1

The Algorithm [Lip87]

1. Initialize all weights to random values. Weight initialization is discussed in Section

2.5.
2. Prepare a data set of input vectors x0, x1, …… xm, with their associated output vectors
d0, d1, …… dm. In our case, we have 729 sets of input/output pairs. Of this, we
reserve 25% (183) for testing (validating) and use the other 546 for training.
3. Present the first input vector for training. Evaluate actual output. Starting at the output
node, work back to the hidden layers and the input layer to adjust the weights, as
described in Section 2.3. Then present second input vector, adjust weights, and repeat
for all 546 examples in the training set.
4. At the end of this training cycle, test the network with the testing data set (183 sets of
input/output pairs). Evaluate Mean Absolute Deviation (Other measures can also be
used). If MAD ceases to improve (See Section 2.7), stop training; else repeat steps 3
and 4 for another training cycle.
N

∑Y
n =1
n − Yˆn
Mean Absolute Deviation (MAD) = (1.30)
N
where N = Testing sample size (N = 183)
Yˆn = Predicted (actual) output of for sample n
Yn = Desired output of sample n

Note that the algorithm described above represents a sequential learning approach. An
alternative to this technique is the batch mode where weights are adjusted at the end of
each training cycle as against weight correction after each presentation of an input vector
(as seen in the sequential approach). An advantage of the sequential approach is its
relative potential to escape from local minima.

2.4 Pre-processing
Neural networks in principle can map raw input values directly into require final output
values. However, in most situations neural networks do not yield satisfactory results
without having the input data pre-processed. In many practical applications, the choice of
pre-processing appears to be one of the most significant factors in determining the
performance of the neural network model [Bis95]. It is often necessary to transform the
input data into some new representation before presenting them to the network.

Each input vector has to be pre-processed before it is passed on to the network.

Sometimes, post-processing of the output values may also be essential. The output values

9
must be post-processed using a post-processing transformation. In the training phase, the
target values must be inverse-transformed to obtain the output values for training.

Standardization
Standardization is one of the most common techniques in pre-processing input data. It is
essentially a linear scaling of the input variables so that different variables having values
which differ by orders of magnitude can be presented in a manner where the relative
importance of the different input variables in determining the output values is maintained.
For example, in our implementation, the MID and the MVED parameters for both the
existing and the new part fall within a range of 40-150, while the Area parameter falls
within a range of 900-9700. The aim is to transform every input variable so that every
variable in the input vector is of the same order of magnitude.

Suppose xi is the sample mean of an input variable x k , σ k2 is the sample variance

with respect to the training set of input data.

M
1
xk =
M
∑x
m =1
m
k (1.31)

1 M m
σ =
2
k ∑
M − 1 m =1
( xk − xk ) 2 (1.32)

where x km (m = 1,…,M) is the value of variable x k in the mth input vector of the training
set, and M is the total number of input vectors in the training set.

Each instance of the input variable is transformed as follows:

x kk − x k
xˆ km = (1.33)
σk

The re-scaled variable x̂ k , associated with x k in the training set, has a zero mean and
unit standard deviation. Note that each variable is transformed independently. This
ensures that all of the input and output variables are of the order unity, and hence, the
weights would also be restricted to order unity. For the first training iteration, weights can
be randomly initialized restricting them to order unity. In the absence of pre-processing
(and, in some cases, post-processing), weights may have differing orders of magnitude.

2.5 Weight Initialization

Most of the initialization techniques aim at setting weights to randomly chosen small
values. This is required in order to avoid symmetries in the network. The sigmoid
function outputs a value very close to zero or one (in other words, the function is
saturated) if it receives an input that lies outside [-3,3]. Therefore, the initial weights
should be small so that the sigmoid transfer functions are not driven into their saturation
regions. However, if weights are too small, the sigmoidal functions will be approximately

10
linear. As a result, nonlinearity in the network will be lost and training will be slower. It
is desirable to maintain the summed inputs to sigmoidal functions within order unity.

Given that the inputs are standardized (rescaled so as to have a zero mean and unit
standard deviation), initial weights can be generated from a Gaussian distribution with a
zero mean and a variance σ2 α 1/d, where d is the number of nodes in the layer where the
connections originate [Bis95]. Suppose there are seven input neurons feeding into the
first hidden layer and thirteen nodes in the hidden layer feeding into the output layer.
Weights associated with arcs that connect nodes in the input layer to nodes in the hidden
layer are generated with a standard deviation σ α 1/√7, while weights associated with
arcs that connect nodes in the hidden layer to nodes in the output layer are generated with
a standard deviation σ α 1/√13. In our implementation, σ = 1/√d.

2.6 Adaptive Learning

Setting a value for the learning rate is essentially a trade-off between speed of
convergence and the ability to closely approximate the gradient path. When ε is small,
convergence will be slow due to a large number of update steps needed to reach a local
minima. With a large ε, convergence is fast initially, but will induce oscillations and the
error function will not reach a minimum [Lin97]. One possible strategy [Has95] is to use
a large step size when the iteration is far from a minimum and a decreasing step size
when the iteration approaches a minimum. However, this induces oscillations around the
region where there is a decrement in the learning rate. Therefore, in practice, a constant
value for the learning rate is used as this generally leads to better results even though the
guarantee of convergence is lost [Bis95].

Another heuristic [Has95] to accelerate learning is to use learning rates, specific for
each node, which are proportional to the number of nodes that feed in. If there are seven
nodes in the input layer and thirteen nodes in the hidden layer, the learning rate for each
node in the hidden layer is inversely proportional to seven, while that for each node in the
output layer is inversely proportional to thirteen.

2.7 Cross-Validation
When the performance of a network is determined by evaluating the error function with
respect to an independent testing (validating) data, it is observed that the validation error
decreases monotonically to a minimum but then starts to increase. In general, multiple
local minima may also exist in the validation error curves. This is attributed to excessive
training, when training with noisy data, leading to over-fitting. Therefore, training is
continued as long as performance on the validation set keeps improving. The set of
weights that yields the least MAD is retained. This method of partial training may lead to
a better generalization.

3. Preliminary Results
This section presents some of the preliminary results and observations for the proposed
model for design similarity prediction. As mentioned earlier, there is no methodology to
select the number of hidden layers, number of nodes in each hidden layer, connectivity,

11
and a learning algorithm. This necessitates constructing networks with different values
for the model parameters to find the optimal configuration.

For training and testing, we created 27 designs and generated each design’s best
fixture using FixtureNet. Then, we evaluated the usefulness of each fixture on itself and
all other designs. This yielded 729 (272) pairs of designs, each with a relative usefulness
metric. Note that, a fixture yields a relative usefulness metric of 1 with its own design.

Feed-forward networks with sigmoid activation functions for hidden and output
layers are considered. Learning is by the Least Mean Square based Back Propagation
Algorithm described in Section 2.3. The learning rate for a network is represented by ε′.
However, the learning rate for each node is determined by the heuristic [Has95] described
in Section 2.6. For example, for a three-layered network with number of hidden nodes =
13, ε′ = 0.07, learning rate for the hidden layer = ε′ /7 = 0.07/7 = 0.01, while the learning
rate for the output layer = ε′ /13 = 0.07/13 = 0.0054.

When ε′ is set at a relatively small value, the algorithm gets caught in local minima
and does not approach the global minimum. Oscillations are observed as the learning rate
is increased (See Figure 2). An appropriate range of learning rate values is determined
and the network is tested for different values of learning rate in this range. In our
implementation, the momentum factor is maintained at η = 0.9 for all experiments.

Figure 2. Oscillations are observed when ε′ = 0.5 for a three-layered network with
seven nodes in the hidden layer.

Experiments are conducted for networks with one and two hidden layers, while the
number of hidden nodes is set at either seven or thirteen (including the bias node).
Neglecting the bias nodes in the hidden and input layers, a network with seven nodes in
the hidden layer has equal number of hidden nodes and input nodes, while a network with
thirteen nodes in the hidden layer has two hidden nodes for every input node. The results

12
are tabulated in Table 1. A configuration is specified by the number of hidden layers n1,
number of hidden nodes n2, and a learning rate ε′. f(n1, n2, ε′) is the minimum MAD
observed in 5000 training cycles.

ε′ f(1,7, ε′ ) f(1,13, ε′ ) f(2,7, ε′ ) f(2,13, ε′ )

0.01 0.217768
0.04 0.219189 0.211462 0.215588 0.210713
0.05 0.214961 0.213404 0.210757
0.07 0.209749 0.21449 0.209364 0.216243
0.09 0.210894 0.217342 0.206576 0.207592
0.1 0.216591 0.218244 0.210914 0.21443
0.2 0.208051 0.21481
0.3 0.207501 0.215314 0.207629
0.5 0.216103
Table 1. Network performance (MAD) Versus Network Structure

A network with two layers and seven nodes in the hidden layer is identified as the
optimal network model and the associated minimum MAD is 0.206576. However, these
experiments are not exhaustive. By further varying learning rates and momentum
parameters, we might arrive at a better network structure. Note that there is little or no
improvement with increase in the number of hidden layers or the number of nodes within
each hidden layer.

A comparison of the neural network model performance against the performance of

other measures based on MID, Area and MVED (obtained from the approach described
in Section 2.1) is presented below. Table 2 shows the correlation coefficient between the
design similarity measure against the desired relative metric for measures based on MID,
Area, MVED and the Neural net. The Neural network-based design similarity measure
performs better than the other models and exhibits a high correlation.

MID Area MVED Neural Network

Correlation 0.3593 0.3432 0.1558 0.6001
Coefficient
Table 2. Correlation Coefficients for various design similarity models

For ten new designs, we compared the best existing fixture identified by different
similarity measures to the best generatively designed fixture (See Table 3). An
observation of the relative metric (ratio of usefulness for the existing fixture of the most
similar part to the usefulness of the generatively designed fixture) of these identified
fixtures indicates that the neural network model, on an average, identifies a fixture with a
higher usefulness metric. In other words, the average relative metric for the best existing
fixture identified by the different measures is highest for fixtures identified by the neural
network model. Also, the design similarity measure based on the neural network model
identifies a better fixture in nine out of the ten cases.

13
 Usefulness Metric for Most Similar Part Relative Metric of Most Similar Part
New MID Area MVED Neural Generative MID Area MVED Neural
Part (Metric) (Metric) Net Fixture (Relative (Relative (Relative Net
Number Metric Metric) Metric) Metric) (Relative
Metric)
1 7.37 0 11.399 14.629 69.292 0.1064 0 0.1645 0.2111

2 20.975 0 21.617 22.344 69.717 0.3009 0 0.3101 0.3205

3 0 0 0 7.832 13.229 0 0 0 0.592

4 0 0 0 16.148 33.773 0 0 0 0.4781

5 0 15.595 15.584 28.454 33.549 0 0.4648 0.4645 0.8481

6 0 0 0 0.695 23.366 0 0 0 0.0297

7 3.768 0 0 0 36.504 0.1032 0 0 0

8 13.977 13.977 13.977 16.726 16.726 0.8356 0.8356 0.8356 1

9 15.426 0 21.127 21.127 28.253 0.546 0 0.7478 0.7478

10 0 0 0 7.36 8.334 0 0 0 0.8831

Average 0.1892 0.13 0.2523 0.5111

Table 3. Performance of the neural network-based design similarity measure versus measures based on MID,
Area, and MVED in identifying the best available fixture.

14
4. Conclusions
This paper presented a variant fixture planning approach that uses a fixture-based design
similarity measure to find promising fixtures quickly. Thus, this approach avoids
checking the feasibility of each existing fixture. For each promising fixture, the approach
calculates a more precise usefulness metric that describes how well the fixture can hold
the new design.

Design attributes that reflect fixture usefulness were identified. Design similarity
measures based on any of these design attributes have been discussed. The use of neural
networks to represent the mapping between design attributes and fixture usefulness has
been explored. Compared to other measures based on design attributes that reflect fixture
usefulness, the neural network-based design similarity measure exhibits better
consistency and, on average, identifies an existing fixture with a higher usefulness metric.

Future work will continue to integrate the variant fixture planner with a generative
planner that creates a preliminary process plan to build a hybrid variant-generative
process planner. This fixture planning approach also could be applied to fixture planning
in other domains.

5. References
[Bal98] S. Balasubramanian, A. Elinson, J. Herrmann, and D. Nau, "Fixture-Based
Usefulness Measures for Hybrid Process Planning," Proceedings of
the1998 ASME Design Engineering Technical Conference, Atlanta,
September 1998.

[Bis95] Bishop, Christopher M., “Neural Networks for Pattern Recognition,”

Oxford University Press Inc., New York, 1995.

[Bro96] Brost, R.C. and Goldberg, K.Y., A Complete Algorithm for Designing
Planar Fixtures using Modular Components, IEEE Transactions on
Robotics and Automation, Volume 12, Number 1, February 1996.

[Has95] Hassoun, Mohamad H., “Fundamentals of Artificial Neural Networks,”

The MIT Press, Cambridge, Massachussetts, 1995.

[Her97] Herrmann, Jeffrey W., and Gurdip Singh, “Design similarity measures for
process planning and design evaluation,” Technical Report 97-74, Institute
for Systems Research, University of Maryland, College Park, 1997.

[Hua99] Huang, C. –L., Huang, Y. –H., Chang, T. –Y., Chang, S. –H., Chung, C. –
H., Huang, D. –T. and Li, R. –K., “The Construction of Production
Performance Prediction System for Semiconductor Manufacturing with
Artificial Neural Networks,” International Journal of Production
Research, Vol. 37, No. 6, pages 1387-1402, 1999.

15
[Lin97] Lin, Z. –C., and Huang, J. –C., “The Applications of Neural Networks in
Fixture Planning by Pattern Classification,” Journal of Intelligent
Manufacturing, Vol. 8, pages 307-322, 1997.

[Lip87] Lippmann, Richard P., “An Introduction to Computing with Neural Nets,”
IEEE ASSP Magazine, pages 4-22, April, 1987.

[Mul90] Muller, B. and Reinhardt, J., “Physics of Neural Networks, Neural

Networks: An Introduction,” Springer-Verlag, 1990.

[Phi94] Philipoom, P. R. and Rees, L. P., Wiegmann, L., “Using Neural Networks
to Determine Internally-Set Due-Date Assignments for Shop Scheduling,”
Decision Sciences, Vol. 25, pages 825-851, 1994.

[Zhu96] Zhuang, Y. and Goldberg, K. Y., “On the existence of Solutions in

Modular Fixturing”, International Journal of Production Research,
Volume 15, Number 6, pages 646-656, 1996.

16