Source: http://cumincad.architexturez.net/system/files/pdf/01f7.content.pdf

describes ai experiment in cg architectural design and explains agent-based systems.

hundreds of agents simultaneously build, manipulate, and dismantle their environment. can work in collaboration, disjunction, or autonomously.

architectural design commonly describes as ability to see whole picture, organize, collect, juggle, manage, and maintain multiple conflicting goals and values; hierarchical and top-down. agent-based experiment presents alternative design process. agents “move through the design landscape, simultaneously collaborating, building, degenerating, and transforming their world.

pg 63

introduction

“experiment in autonomous self generated architectural design.:
“embedding fragments of intelligence and design sensibility within architectural objects so that they might learn how to search through design possibilities autonomously.”

“The application works by first selecting the agent modules and agent behaviors to be used within the simulation. Once the simulation is started the viewer watches the environment”
non-interactive, passive approach. no opportunity for changing inputs temporally.

KBAI vs. BBAI

Knowledge Based Artificial Intelligence (KBAI) – top-down, model single competencies, depth more important that breadth of search, closed and non-autonomous, info is interpreted through reasoning and planning, empirical knowledge is not aquired within the system because this knowledge is built in from the start.
Behavior based artificial intelligence (BBAI) – evolving multiple competencies at a very low level, exhibit competencies which can describe various ‘design processes’ and internal representations at a different level from KBAI. rudimentary agents which either grow in capability through collaboration, union, or disjunciton, or learn through countless ‘tests’ driven by fitness (Koza 1992). autonomous and open. Agents within these systems find, evolve, and derive their own goals. task-oriented decomposition with task-dependant modules. internal representations of problems, goals, or actions within the agents are usually multiple, redundant, and possibly inconsistent and nonobjective.

pg 64

agent model-emergence: “in a formal system, it is something such as form, which ’emerges’, though not explicitly coded. complex relationships between programmed structure for generation and the potential outputs the beauty of the system evolves not through restricting possibilities but by understanding the extents of potential emergence.”

bounding possibilities
simple example: two lines, only be connected at ends, 4 states or 22 possibilities
also allow them to be connected at midpoints, 32 possible connections.
allow any length or connect at any point, infinite potential

in agent generated system we start with infinite sets and refine solutions through interactions.
“In infinite sets, refinement and emergence become helpful tools to quantize predicable family classes (Dawkins 1988)”

can still be infinite distinct objects but perceivable grouping of families can become the evaluation factor for design evolution. “The offspring (objects) may be infinite in describable appearance, but the families are limited in description and grouping characteristics within the generated set”

pg 65

“The point is not to predict or work towards a larger ordering system which describes standard accepted notions of proportion or ‘conventional beauty’, but o celebrate the possibity of a more chaotic intervention within a simple structure which can ’emerge’ more didactic architectural (formal) relationships.”

“agents perform in a field called the ‘environment'”
“‘user’ can determine the number and type of agents within the system. The agents are then randomly generated to start with an inconsistent and chaotic pool of participants.”

Point agents move through space and can communicate with other agents. Point agents have no formal outcome except through the visualization of the path the agent took during the simulation. Point agents can also ‘carry’ agents with them to build second order agents.”
Polygon agents are three-dimensional objects (consisting of 4 point agents). Polygons move through the environment interacting with other agents.”
Second order agents are objects made by the combination of other agents (i.e. combining multiple polygon agents).”
Third order agents are static derivatives of second order agents. Once generated third order agents cannot be removed even if the parent second order agent is destroyed or moved. Third order agents can be augmented by abstractors and articulators (these are beyond the scope of this paper).”

pg 66

agents have behaviors that allow them to:

  • generate, augment, or add a particular component to another (potentially new) object or itself
  • substract objects, properties, or characteristics from another agent or itelf
  • change the state of objects or agents
  • aggressive agents will try to change as much as it can (multiple attributes) within other agents
  • collaborative agents will look for a best fit when communicating with other agents

pg 67

conclusion

“Agent based modeling can be used as a technique for creative form generation, or to emerge basic architectural principles autonomously (symmetry, tectonics, linearity, grouping, layering).”

“Behavior based investigations can lead to fundamentally new ways of approaching design and creative architectural and generative experiences. This project serves as an introduction to such investigations and as a departure from traditional top-down computational approaches. Future experiments could examine embedding elementary knowledge into architectural agents for use in CAAD software, or more process oriented and flexible form generation and rendering packages.”

pg 69

Agent Generated Software written in C++

pg 70