The advent of search algorithms that are inspired by biological evolution is a story of co-evolution developing
independently in the United States and Germany in the 1950s and 1960s. In both cases, principles of biological
evolution such as natural genetic variation and natural selection were applied to computer programs to solve
parameter optimization (satisficing) problems. In Germany “evolution strategies” were introduced by Rechenberg
and later Schwefel which could be used to solve real world engineering problems such as the design of airfoils
(Back et al., 1991) (Mitchell, 1998). At the same time Holland was developing Genetic Algorithms (GA) which
were attempts at formalizing evolutionary and adaptation processes in nature and applying them to computer
systems (Holland, 1975). Genetic algorithms were an abstraction of organic evolution replacing chromosomes with
bits (strings of ones and zeros) and using the operators: crossover, mutation and inversion to “evolve” populations
based on sexual reproduction. In a GA, populations evolve over time relative to some fitness criteria or objective
function either exogenous or endogenous (Mitchell, 1998). Evolutionary programming was another line of research
which was also developed in the 60s. Here, a finite-state machine representing a given solution is mutated according
to a state-transition diagram to develop more fit individuals (Fogel, Owens, Walsh, 1966) (Mitchell, 1998).
Additional work has been done by Koza (1992) termed genetic programming (GP). Genetic programming uses a
different method than the binary strings Holland uses. GP begins with parse trees (hierarchical node graphs)
consisting of nodes representing various mathematical operators, variables, and logic gates: [ +, –, *, / , A, X]
(Figure 2.1.1). Trees are then randomly cut at a node and inserted into other trees, or subtrees may be removed and
replaced with other subtrees and recombined over many generations relative to a training set on some data. Koza
claims such a system is capable of automated programing and has demonstrated GPs effectiveness within a variety
of domains. GP’s has been compared to standard search algorithms hill climbing and simulated annealing by
O’Reilly and Oppacher (1994) for five problems and has been shown to be equal at best, or less effective.
Koza, Mitchell and Crutchfield (Koza, 1992) (Mitchell, Hraber, Crutchfield, 1993) have experimented with
automated programing using GAs on evolving one-dimensional binary (black, white) cellular automata (CA). White
investigates the development of a cellular automaton to model the spatial structure of urban land use over time
(White, 1993). CAs represent a system where simple rules govern local interactions yet global or emergent behavior
may arise in the absence of any governor or central processor. Mitchell and Crutchfield used GAs to perform
density-classification tasks on CAs such as determining if the initial configuration contains majority black or white
cells. Rules evolved by the GA have been discussed in terms of computational mechanics relating to concepts such
as, “particles” and “particle interactions” (Fig. 2.1.2) (Crutchfield, Mitchell, Das 1996). Schelling applies CAs to
models of human neighborhood segregation (Schelling, 1971).
The preceding discussion illustrates just a few of the many GAs that have been developed. As
Mitchell describes, specific GAs are as numerous as the problems they attempt to solve (Mitchell, 1998) and in this
sense do not represent a universal search algorithm but rather a heuristic approach or process that
is tailored by the scientist / designer. GAs have been widely used not only for discriminative data
analysis but also as generative algorithms. Following the previously mentioned pioneering work of
Karl Sims, GAs have been incorporated by designers in a variety of ways. Artists such as Rooke,
Unemi and Hart followed in Sims’ footsteps using expression based algorithms similar to Koza’s
genetic programing model and many contemporary visual artist have followed this earlier work (see
Romero, 2008). 3D GAs have also been developed with a variety of approaches such as geometric
(lattice) deformation by Wattabe and sequences of polygonal operators by McGuire (Romero, 2008).
Latham and Todd developed the PC mutator system at IBM UK’s Scientific Centre with individual
projects as well as commercially available software (Romero 2008). Hemberg and the Emergent Design
Group (EDG) at MIT developed Genr8 in 2001 which is a GA plug in for the modeling and animation
software Maya (Hemberg, 2006, 2007). EDG is an interdisciplinary group that attempts to bridge
computer science and architecture in the nascent field of generative architecture.
Generative architecture grew out of the research in Cybernetics of the 1950s. Frazer developed the
Evolutionary Digital Design Process at the Architecture Association (AA) as part of his work in
morphogenesis in the mid 90s which has been influential. Frazer describes parametric design, or
parameticism, as a process driven methodology that has the potential to address larger (global)
environmental concerns. Frazer taught with Gordon Pask who introduced cybernetics into
architectural theory in the 60s through Stafford Beer and Von Neumann’s “Universal Constructors”
which Pask developed as 3D building block prototypes (Frazer, 1995, 1997). Other research
architects such as Coates cite Turing’s early work in reaction diffusion processes (Turing, 1952)
and Lindenmeyer’s L-systems (Lindenmeyer, 1968) to being influential in his work on shape grammars
to breed structures that respond to physical constraints in the environment such as light and wind
(Coates, 1999). Coates utilizes a form of GP that uses L- systems as parse trees which evolve
relative to an objective function. Combining GP and L-systems is a parsimonious approach to
exploring what Coates terms, “spatial morphogenesis” (Coates, 1999) 1. Weinstock is also an
important contributor to generative architecture in terms of emergence and ideas relating to
evolution and form in architecture, drawing from a macro view of geology, biology and urban forms
(Weinstock, 2010). Precedent in architecture and evolutionary design is investigated by Zarzar and
earlier by Gero. Zarzar develops the notion of d- genes and focuses on the architecture of
Corbusier and Calatrava, analyzing form, structure and the evolution of design principles such as
Corbusier’s five points of modern architecture (Zarzar, 2003). Gero does research into the
architecture of Frank Lloyd Wright (FLLW) and the paintings of Mondrain. Gero develops a GA to
abstract rules from a number of FLLW window compositions and Mondrain paintings and then uses the
GA to develop new designs that are a hybrid of both architect and artist (Fig. 2.1.3) (Gero, 1998).
Patrick Schumacher writes in “The Autopoiesis of Architecture” that “Beyond such obvious surface
features one can identify a series of new concepts and methods that are so different from the
repertoire of both traditional and modern architecture that one is justified in speaking of the
emergence of a new paradigm within architecture. New design tools play a crucial part in making this possible, establishing a whole new design process and methodology” (Schumacher, 2012). Trends
in research are towards biomimetic approaches combining evolutionary computational methods with morphogenetic processes
inspired by nature, where form is generated by computer technology, incorporating the rules and constraints of
fabrication (Menges, 2012, 2013). GAs tools are becoming commercially available for practitioners such as the
Genr8 Maya plugin mentioned above as well as the Galapagos plugin for Rhino-Grasshopper. Galapagos and
Grasshopper offer a visual interface for architects that allows for GAs to be experimented with in a variety of ways.
Galapagos has been used to optimize spatial adjacencies for complex building programs (Boon, Griffin et al. 2015).
This project optimizes a three dimensional layout for 50 programmatic spaces, essentially creating a bubble diagram
that an architect may then use for schematic design. Galapagos has been used for daylighting and shading studies
(Gonzales, Fiorito, 2015) as well as to find novel solutions to structural problems (Danhaive, 2015). Galapagos has
also been used to generate new fractal forms for urban environments using random curds and cellular automata
(Devetakovic, 2015).
The brief background above shows the coming together of architecture and computer science towards a new
evolutionary design approach that combines the creativity of the designer with the computational power of
computers to simulate evolutionary processes. This process draws heavily from nature and has been termed
biomimetic design as well as generative design. One abstract idea seen again and again in biomimetic design is
fractal geometry. The use of L-systems and fractals in architecture, urban planning and biomimetic design will be
discussed next in section 2.2.
