• Functional Scalability through Generative Representations ...

  • FileName: hornby_epb04.pdf [read-online]
    • Abstract: We argue that generative representations, those. which are capable of ... To support this argument we compare a generative and non-generative representation on a ...

Download the ebook

Functional Scalability through Generative
Representations: the Evolution of Table Designs
Gregory S. Hornby
QSS Group Inc., NASA Ames Research Center
Mail Stop 269-3, Moffett Field, CA 94035-1000
[email protected]
One of the main limitations for the functional scalability of automated design systems is the
representation used for encoding designs. We argue that generative representations, those
which are capable of reusing elements of the encoded design in the translation to the actual
artifact, are better suited for automated design because reuse of building blocks captures
some design dependencies and improves the ability to make large changes in design space.
To support this argument we compare a generative and non-generative representation on a
table design problem and find that designs evolved with the generative representation have
higher fitness and a more regular structure. Additionally the generative representation was
found to better capture the height dependency between table legs and also produced a wider
range of table designs.
Key words: Generative representation, Representation, Genetic algorithms, Evolutionary
design, Lindenmayer Systems (L-systems)
1 Introduction
Computer automated design systems have been used to design a variety of differ-
ent types of artifacts such as antennas (Lohn et al. , in press), flywheels, load cells
(Robinson et al. , 1999), trusses (Michalewicz et al. , 1996), robots (Lipson & Pol-
lack, 2000), and more (Bentley, 1999) (Bentley & Corne, 2001). While they have
been successful at producing simple, albeit novel artifacts, a concern with these
systems is how well their search ability will scale to more complex design spaces
(Drexler, 1989) (Pollack et al. , 2001). To improve functional scalability we can
look at those fields which regularly create large, complex artifacts. In engineering
and software development, complex artifacts are achieved by exploiting the prin-
ciples of regularity, modularity and hierarchy (Ulrich & Tung, 1991) (Huang &
Preprint submitted to Environment and Planning B 1 November 2004
Kusiak, 1998) (Meyer, 1988), which can be summarized as the hierarchical reuse
of building blocks.
While the optimization algorithm can affect the degree of reuse in a design, the
ability to create structures which reuse building blocks is limited by the ability
of the representation to encode them. For example, in optimizing the dimensions
on a blueprint, the design system can only produce designs that fall in the pre-
specified parameter space. A limitation with this type of representation is that no
modification of the search algorithm can affect the degree of reuse in resulting
designs, nor is the hierarchical construction of building blocks possible. Thus the
ability to automatically generate structures which have a reuse of subassemblies is
strongly dependent on the representation used by the design system.
The different types of representations for computer-automated design systems can
be classified by how they encode designs. First, designs can be split into parameter-
izations or open-ended representations. Parameterizations consist of a set of values
for the dimensions of a pre-defined blueprint and open-ended representations are
those in which the topology of a design is changeable. Since one of the goals of au-
tomated design systems is to achieve truly novel artifacts, we focus on open-ended
representations because it is difficult for a parameterization to achieve a type of de-
sign that was not conceived of by its creators. A fundamental distinction between
open-ended representations is whether they are non-generative or generative. With
a non-generative representation each representational element of an encoded de-
sign can map to at most one element in a designed artifact. The two subcategories
of non-generative representations are direct and indirect representations. With a
direct representation, the encoded design is a blueprint in which elements can be
added/removed in addition to changing their parameters, and with an indirect rep-
resentation there is a translation or construction process in going from the encoding
to the blueprint. A generative representation is one in which an encoded design
can reuse elements of its encoding in the translation to an actual design. The two
subcategories of generative representations are implicit and explicit. Implicit, gen-
erative representations consist of a set of rules that implicitly specify a shape, such
as through an iterative construction process similar to a cellular automata and ex-
plicit, generative representations are a procedural approach in which a design is
explicitly represented by an algorithm for constructing it.
Previously we have argued that the advantage of generative representations over
non-generative representations is that they incorporate useful bias into their struc-
ture (Hornby & Pollack, 2002). Here we refine and extend these arguments in two
ways and support these claims by comparing optimization with a generative rep-
resentation against optimization with a non-generative representation on our table
design problem (Hornby & Pollack, 2001). First we argue that optimization per-
forms better with a generative representation because generative representations are
better at capturing some types of design dependencies than non-generative repre-
sentations. We support this argument by showing that evolution with the generative
representation was better able to create multi-legged tables. Second we claim that
generative representations are more conducive to changes thereby improving the
ability of the search algorithm to move about in the design space. This is demon-
strated by the greater variety of styles of tables produced with the generative repre-
The rest of the paper is organized as follows. First we present our arguments for
the advantages of generative representations followed by a review of different auto-
mated design systems. Next we describe the parts of our evolutionary design system
and then the results of our experiments in evolving table designs. Finally we close
with a discussion of our findings and a summary of this paper.
2 Argument for Generative Representations
As the complexity and number of parts in a design grows, the functional scalability
of non-generative representations is limited by their weakness in handling the in-
creasing number of design dependencies and the exponential growth in the size of
the search space. In the first case, as designs become more complex dependencies
develop between parts of a design such that changing a property of one part requires
the simultaneous change in another part of the design. For example, if the length of
a table leg is changed, then all of the other legs must be changed or the table will
become unbalanced. Second, as a design grows in the number of parts the expected
distance (in number of parts) between a starting design and the desired optimized
design increases. Conversely, changing a single part makes proportionately smaller
and smaller moves towards the desired design. One consequence of this is that as
designs increase in the number of parts, search algorithms require more steps to
find a good solution. Increasing the size of variation (by changing more parts at a
time) is not a solution because as the amount of variation is increased, the probabil-
ity of the variation being advantageous decreases. Non-generative representations
are not well suited to handling these increases in size and complexity because their
language for representing designs is static.
The advantage of generative representations comes from their ability to reuse pre-
viously discovered building blocks. First, reuse of elements of an encoded design
allows a generative representation to capture design dependencies by giving it the
ability to make coordinated changes in several parts of a design simultaneously.
For example, if all the legs of a table design are a reuse of the same compo-
nent, then changing the length of that component will change the length of all
table-legs simultaneously. Secondly, navigation of large design spaces is improved
through the ability to manipulate assemblies of components as units. For exam-
ple, if adding/removing an assembly of m parts would make a design better, this
would require the manipulation of m elements of a design encoded with a non-
generative representation. With a generative representation the ability to reuse pre-
viously discovered assemblies of parts enables the addition/deletion of groups of
parts through the addition/deletion/modification of symbols which represent groups
of parts. Here the ability to hierarchically create and reuse building blocks acts as
a scaling of knowledge through the scaling of the unit of variation.
3 Review of Automated Design Representations
To assist in our review of design representations, we first define some of their prop-
erties. Previously we used the metaphor of design representations as a kind of com-
puter programming language to define the following features of design representa-
tions (Hornby & Pollack, 2002):
• Combination: The ability to hierarchically create more powerful expressions
from simpler ones. While the subroutines of GLib (Angeline & Pollack, 1994)
and genetic programming (GP) (Koza, 1992) allow explicit combinations of ex-
pressions, combination is not fully enabled by mere adjacency or proximity in
the strings utilized by typical representations in genetic algorithms.
• Control-flow: All programming languages have some form of control of exe-
cution which permits the conditional and repetitive use of structures. Two types
of control-flow are conditionals and iterative expressions. Conditionals can be
implemented with an if-statement, as in GP, or a rule which governs the next
state in a cellular automata. Iteration is a looping ability, such as the for-loop
in C/C++ programs.
• Abstraction: This consists of the ability to encapsulate a group of expressions
in the language and label them, enabling them to be manipulated/referenced as a
unit, and the ability to pass parameters to procedures. An example of abstraction
is the automatically defined functions (ADFs) of GP.
An open-ended representation is generative if it has reuse (which can be through
either iteration, procedure labels, or both), otherwise it is non-generative. In the rest
of this section we review different design representations for both non-generative
and generative representations.
The two classes of non-generative representations are direct and indirect. Direct,
non-generative representations typically use a fixed-size data structure for speci-
fying the existence of material at a given location, such as with two-dimensional
arrays (Kane & Schoenauer, 1996) (Baron et al. , 1999). Indirect, non-generative
representations tend to be variable length assembly procedures which specify the
assembly of an object (Bentley, 1996) (Funes & Pollack, 1998). An additional layer
of indirection is used in those representations in which an object is encoded as a
derivation tree for a grammar, with the resulting assembly procedure specifying the
assembly of the artifact (Roston, 1994).
One type of grammar for design is Stiny’s shape grammars (Stiny, 1980). The gram-
matical rules of a shape grammar specify a transformation from one shape to an-
other. Optimization with shape grammars consists of producing a derivation tree
for the grammar, with a typical approach using simulated annealing to iteratively
modify a design (Shea et al. , 1997) (Agarwal & Cagan, 1998). While automated
design using shape grammars is an example of an indirect representation (since no
element of the derivation tree is used more than once) it could be extended to a gen-
erative representation by allowing new rules to be added to the grammar through
the combination of existing rules.
Most implicit, generative representations consist of a starting shape and a set of
rules for iteratively transforming the design. One of the earliest such examples is
Frazer’s work using shape transformation rules to rotate/stretch/grow/shear a start-
ing shape (Frazer, 1995). Similar to Frazer’s work is that of de Garis’ augmented
cellular automata (CA) (de Garis, 1992), in which each cell in the CA maintains
the number of neighbor cells in the ON state in each of the four directions, North,
East, West and South. More standard CAs are used in (Bentley & Kumar, 1999)
for creating two-dimensional tessellating tile patterns and for creating patterns of
cells in an isospatial grid (Bonabeau et al. , 2000). Rather than working in a grid,
Shape Feature Generating Process (SFGP) grows designs by optimizing rules for
the division of dots (metaphors for a cell) on the surface of a shape (Taura et al. ,
1998) (Taura & Nagasaka, 1999). After development is complete, the final shape
is formed by creating an outer surface using the density of dots to determine the
distance from the initial shape. A more biologically based model is Eggenberger’s
method of growing three-dimensional shapes from an artificial genome using an
artificial chemistry (Eggenberger, 1997). Designs are encoded in a linear string
which consists of regulatory genes, for switching other genes in the genome, and
structural genes, which encode for specific chemicals.
Explicit, generative representations consist of an algorithm for constructing a de-
sign and are typically implemented as a type of grammar. With Todd and Latham’s
Mutator, structures are defined by an expression in a geometrical construction lan-
guage that specifies the shape, shape transformations, number of repetitions of a
shape and angles between shapes. Initially only the evolution of parameters was
possible (Todd & Latham, 1992), and then in later work the ability to change
the grammar was added (Todd & Latham, 1999). Broughton, Coates and Jack-
son (Broughton et al. , 1997) (Coates et al. , 1999) use a Lindenmayer system
(L-system) as the representation for evolving shapes in an isospatial grid and the
Emergent Design group has used L-systems with attractors/repellers for evolving
curved surfaces in a three-dimensional space (Testa et al. , 2000) (Hemberg et al.
, 2001). Rosenman (Rosenman, 1996) (Rosenman, 1997) describes a hierarchical
grammar for building two-dimensional, grid-based, floor plans which uses multiple
evolutionary runs to evolve different levels of the design. Instead of constructing a
final design shape from a number of simpler shapes, the work of Husbands et. al.
(Husbands et al. , 1996) and Nishino et. al. (Nishino et al. , 2001) combines su-
Table 1
Properties of the different design representations.
Combination Control Flow Abstraction
System Iter. Cond. Labels Param.
Direct Non-generative
(Baron et al. , 1999) no no no no no
(Kane & Schoenauer, 1995) no no no no no
Shape grammar systems no no no no no
Indirect Non-generative
(Bentley, 1996) yes no no no no
(Bentley & Kumar, 1999), explicit yes no no no no
(Funes & Pollack, 1998) yes no no no no
Genetic Design (Roston, 1994) yes no no no no
GENRE, non-generative (section 4.3) yes no no no no
Implicit Generative
(Bentley & Kumar, 1999), implicit no yes yes no no
(Bonabeau et al. , 2000) no yes yes no no
(de Garis, 1992) no yes yes no no
(Eggenberger, 1997) no yes yes no yes
(Frazer, 1995) no yes yes no no
(Taura et al. , 1998) no yes yes no no
Explicit Generative
(Broughton et al. , 1997) yes no no yes no
Emergent Design Group yes no no yes no
GENRE, generative (section 4.2) yes yes yes yes yes
Mutator (Todd & Latham, 1992) yes yes no no no
(Rosenman, 1997) yes no no yes no
Superquadrics yes no no yes no
perquadric modeling primitives (Barr, 1981) with constructive solid geometry as a
kind of genetic program for transforming a starting shape.
The evolutionary design systems described in this section are listed in table 1. In
this table the representations are grouped by category and for each representation it
is stated whether or not it has the property of combination, iteration, conditionals,
labels or parameters.
4 Evolutionary Design System
The computer-automated design system used to create designs is called GENRE
and it consists of the design constructor, the compiler for the generative represen-
tation, the fitness function for evaluating designs and the evolutionary algorithm.
The evolutionary algorithm evolves encoded designs using the fitness function to
score designs. To allow for comparing non-generative and generative representa-
tions, GENRE has both a non-generative and a generative representation for encod-
ing designs. The non-generative representation encodes designs indirectly using a
sequence of construction commands, called an assembly procedure, for building
a design with the design constructor. The generative representation is based on
Lindenmayer systems (Lindenmayer, 1968), which are compiled into an assembly
procedure that is used to build the design. By using an indirect, non-generative rep-
resentation and an explicit, generative representation, these two representations can
be applied to different design substrates by changing only the set of construction
commands and the design constructor. The following subsections describe each of
these parts.
4.1 Design Constructor
The design constructor starts with a single cube with which it creates a more com-
plex object by executing the instructions of an assembly procedure. Commands in
the command set act on the local state – which consists of the current position and
orientation – and are listed in table 2. The commands ‘[’ and ‘]’ push and pop
the current state to and from a stack. Forward adds cubes in the positive X direc-
tion of the local state and back adds cubes in the negative X direction. In addition
to adding cubes forward and back also change the current position. The com-
mands left/right/up/down/clockwise/counter-clockwise rotate
the current heading about the appropriate axis in units of 90 ◦ .
The images in figure 1 show intermediate stages in the construction of an object
from the assembly procedure: forward(2) right(1) forward(1) up(1)
forward(3). Initially there is a single cube in the design space, figure 1.a. Af-
ter executing the command forward(2), two cubes are added to the first, fig-
ure 1.b. The image in figure 1.c shows the design after executing right(1)
forward(1), which turns the current orientation 90 ◦ to the right and then adds a
Table 2
Design language for constructing tables.
Command Description
[] Push/pop state to stack.
forward(n) Move and add cubes in the local, positive X direction n
back(n) Move and add cubes in the local, negative X direction n
clockwise(n) Rotate local heading n × 90◦ about the X axis.
counter-clockwise(n) Rotate local heading n × −90◦ about the X axis.
left(n) Rotate local heading n × 90◦ about the Y axis.
right(n) Rotate local heading n × −90◦ about the Y axis.
up(n) Rotate local heading n × 90◦ about the Z axis.
down(n) Rotate local heading n × −90◦ about the Z axis.
cube in the current forward direction. After executing up(1) forward(3), the
final object is shown in figure 1.d. In building an object, if the constructor is asked
to place a cube where one already exists, it ignores the existing cube but updates
its location as if it had placed this cube and then continues executing the assembly
4.2 Generative Representation
The generative representation for each design is based on a grammatical rewrit-
ing system called Lindenmayer Systems (L-systems) (Lindenmayer, 1968). The
class of L-systems used as the encoding for designs in this work is parametric L-
systems. Production rules for a parametric L-system consist of a rule-head, which
is the symbol to be replaced, followed by a number of condition-successor pairs.
The condition is a boolean expression on the parameters to the production-rule,
and the successor consists of a sequence of characters that replace the rule-head.
Rule-head symbols are re-written by testing each of their conditions sequentially,
and replacing the rule-head symbol with the successor of the first condition that suc-
ceeds. Thus the parametric L-system has the properties of combination, abstraction,
naming of compound procedures, formal parameters to the procedures, and condi-
tionals. To handle iteration, a looping ability is added that replicates the symbols
enclosed with parenthesis similar to for loops in computer programs: { block }(n)
repeats the enclosed block of symbols n times. For example {abc}(3) translates to,
(a) (b)
(c) (d)
Fig. 1. Steps in building an object.
Compiling an L-system into an assembly procedure consists of starting with a sin-
gle symbol and then iteratively applying the production rules in parallel to all com-
mands in an assembly procedure. The following is an example design encoded with
the generative representation and the construction language of table 2. It consists
of two productions with each production containing one condition-successor pair:
P 0(n0 ) : n0 > 1.0 → [ P 1(n0 ∗ 1.5) ] up(1) f orward(3) down(1) P 0(n0 − 1)
P 1(n0 ) : n0 > 1.0 → { [ f orward(n0 ) ] lef t(1) }(4)
Compiling this design encoding starting with the symbol P0(4) produces the fol-
lowing sequence of strings,
1. P0(4)
2. [ P1(6) ] up(1) forward(3) down(1) P0(3)
3. [ { [ forward(6) ] left(1) }(4) ] up(1) forward(3) down(1) [ P1(4.5) ] up(1)
forward(3) down(1) P0(2)
4. [ { [ forward(6) ] left(1) }(4) ] up(1) forward(3) down(1) [ { [ forward(4.5)
] left(1) }(4) ] up(1) forward(3) down(1) [ P1(3) ] up(1) forward(3) down(1)
5. [ { [ forward(6) ] left(1) }(4) ] up(1) forward(3) down(1) [ { [ forward(4.5)
] left(1) }(4) ] up(1) forward(3) down(1) [ { [ forward(3) ] left(1) }(4) ]
up(1) forward(3) down(1)
6. [ [ forward(6) ] left(1) [ forward(6) ] left(1) [ forward(6) ] left(1) [ for-
ward(6) ] left(1) ] up(1) forward(3) down(1) [ [ forward(4.5) ] left(1) [ for-
ward(4.5) ] left(1) [ forward(4.5) ] left(1) [ forward(4.5) ] left(1) ] up(1) for-
ward(3) down(1) [ [ forward(3) ] left(1) [ forward(3) ] left(1) [ forward(3)
] left(1) [ forward(3) ] left(1) ] up(1) forward(3) down(1) forward(3)
This final string is a sequence of construction commands which is then used by the
design constructor to build an object.
4.3 Non-generative Representation
To show the advantages of a generative representation it must be compared against
a non-generative representation. For the non-generative representation each indi-
vidual in the population is an assembly procedure which specifies how to construct
the design. We implement this assembly procedure as a degenerate, parametric L-
system which has only a single rule and no iterative loops or abstraction. Imple-
menting the non-generative representation in the same way as the generative repre-
sentation allows us to use the same evolutionary algorithm and variation operators
so that the only difference between evolutionary runs with the two systems is the
ability to hierarchically reuse elements of encoded designs.
4.4 Evolutionary Algorithm
An evolutionary algorithm is a population-based, search algorithm used in opti-
mization. Search operates by creating an initial population of candidate designs,
called individuals, and then iteratively selecting better individuals to reproduce and
make new designs until the search is done. The evolutionary algorithm and varia-
tion operators used by GENRE are described in detail in (Hornby, 2003), here we
give an overview of the system.
The evolutionary algorithm used to evolve designs is the canonical generational
EA with specialized variation operators. Each individual in the initial population is
an L-system with a random set of production rules. After all individuals have been
evaluated, better individuals are selected as parents to create a new population. New
individuals are created through applying mutation or recombination (chosen with
equal probability) to individuals selected as parents. Mutation takes a single indi-
vidual as a parent, makes a copy of it and then makes a small random change to this
child copy. Some of the changes that can occur are inserting a small sequence of
random commands, deleting a small sequence of commands, changing the param-
eters to a command, changing the parameters of a conditional, and encapsulating
a sequence of commands into its own production rule. Recombination takes two
individuals as parents, makes a copy of the first individual and then inserts a small
part of the second parent into this child. Examples of some of the insertions that can
be done are replacing a subsequence of commands in the child with a subsequence
of commands from the second parent, replacing one of the child’s successors with
one from the second parent, and replacing a complete condition-successor pair in
the child with one from the second parent. This process of evaluation, selection,
and reproduction is then repeated for a fixed number of generations.
To reduce the frequency of variations that do not result in a change in the resulting
design, data is kept for compiled L-systems on which production rules and suc-
cessors were used, as well as the value range for each parameter. This compilation
history is used so that variation operators are then applied only to those produc-
tion rules that were used in compiling and so that mutated of condition values stay
within the value range of the parameter being compared against.
4.5 Fitness Function
Once an assembly procedure is executed the resulting structure is evaluated by
a pre-specified fitness function. First, the design simulator determines whether or
not the object is balanced or will fall over. Objects that are not balanced are given a
fitness of zero, otherwise an object is given a score based on several easy to compute
attributes of a table. These attributes are the height at which it holds objects, the
amount of surface area available, how stable it is, and the amount of material it
is made of. Height is simply the number of cubes above the ground and surface
area is the number of cubes at the maximum height. Rather than measuring actual
structural integrity, a rough measure of stability is calculated by using the volume
of the table from summing the area at each layer of the table. The measurement
of amount of material used includes only those cubes not on the surface and is
necessary since maximizing height, surface area and stability typically result in
table designs that are solid volumes.
fheight = the height of the highest cube, Ymax .
fsurf ace = the number of cubes at Ymax .
fstability = farea (y)
farea (y) = area in the convex hull at height y.
fexcess = number of cubes not on the surface.
For these experiments we combine these measures into a single function,
fitness = fheight × fsurf ace × fstability /fexcess (1)
5 Results
To compare the generative representation against the non-generative representa-
tion the evolutionary algorithm was configured to run for two thousand generations
using a population size of two hundred. In this section we present results compar-
ing fitness and evolvability of designs produced with both representations and also
show tables constructed from evolved designs.
generative 6000 generative - L-sys
non-generative generative - assembly proc.
average fitness
average length
3e+06 4000
0 0
0 250 500 750 1000 1250 1500 1750 2000 0 250 500 750 1000 1250 1500 1750 2000
generation generation
(a) (b)
Fig. 2. Graphs comparing (a) average fitness and (b) average length of the design encodings
and assembly procedure.
The graph in figure 2.a contains a comparison of the fitness of the best individ-
uals evolved with the non-generative representation against the best individuals
evolved with the generative representation, averaged over fifty trials. With the non-
generative representation, fitness improved rapidly over the first 300 generations,
then quickly leveled off, improving by less than 25% over the last 1700 generations.
Fitness increased faster with the generative representation, and the rate of increase
in fitness did not slow as quickly as with the non-generative representation. The
final results are an average best fitness of 1826158 with the non-generative repre-
sentation and 4938144 with the generative representation.
(a) (b)
(c) (d)
(e) (f)
(g) (h)
Fig. 3. The four best tables evolved with: (a)-(d), the non-generative representation; and
(e)-(h), the generative representation.
Non-Generative Representation (1 mutation) Generative Representation (1 mutation)
1e+06 1e+06
100000 100000
10000 10000
1000 1000
100 100
10 10
1 1
-300000 0 300000 600000 -300000 0 300000 600000
fitness change fitness change
(a) (b)
Fig. 4. Graph of the number of offspring (y-axis, log scale) that had a given fitness differ-
ential (x-axis) from their parent.
Non-Generative Representation Generative Representation
0.008 0.008
0.0072 0.0072
prob. of improvement
prob. of improvement
0.0064 0.0064
0.0056 0.0056
0.0048 0.0048
0.004 0.004
0.0032 0.0032
0.0024 0.0024
0.0016 0.0016
0.0008 0.0008
0 0
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
command difference between assembly procedures command difference between assembly procedures
(a) (b)
Fig. 5. Probability of success (child is more fit than parent) comparison between
non-generative and generative representations, for ranges 1-50, 51-100, 101-150, ...
In addition to having higher fitness, tables evolved with the generative represen-
tation show some symmetries and regularities whereas tables evolved with the
non-generative representation tended to be irregular, figure 3. These regularities are
most likely a result of the generative representation’s ability to reuse elements. The
average amount of reuse with the generative representation can be calculated from
the average length of the encoded design and the average length of the assembly
procedure it compiles to. From the graph in figure 2.b it can be seen that these val-
ues are approximately 310 and 5100 respectively, which leads to an average reuse
of just over sixteen elements.
To determine if the generative representation produces encodings that are more
conducive to evolution we compare the success rate of the mutation operator. The
graphs in figure 4 plot the number of offspring that fall at a given fitness differential
from the parent design. These graphs show that the vast majority of mutations to
a design produced little or no change in fitness. While most of the remaining mu-
tations produced a negative change in fitness with both representations, there are
(a) (b)
(c) (d)
Fig. 6. Evolved tables shown both in simulation (left) and reality (right).
more positive changes to fitness with the generative representation, especially large
positive changes. A plot of the success rate under mutation (a child has higher fit-
ness than its parent) is shown in the graphs in figure 5. On this graph it can be seen
that mutations that produce a large change in the assembly procedure have a greater
rate of success with the generative representation than with the non-generative rep-
resentation. These two sets of graphs show that designs encoded with the generative
representation are more evolvable than those encoded with the non-generative rep-
Previous work has shown the successful transfer from design to reality of static
objects (Funes & Pollack, 1998) and robots (Lipson & Pollack, 2000). Similarly,
designs produced with this system have also been successfully transferred to the

Use: 0.1358