Sources

	Modeling of 3D scenes often requires plants and trees. 3D modeling of plants is often a tedious process that is not based 
on any real-life process. General plant modeling has usually been accomplished with L-Systems which are fractal-like systems which

recursively define plant growth and allow the plants to iteratively grow more and more complex. Traditionally L-Systems have been 

implemented in 2 dimensions. However, some recent papers have suggested using 3D L-Systems to model 3-dimensional plants.
	An L-System consists of a set of substitution rules (usually just one) which represent the plant. The rules consist of 
commands similar to the LOGO language and places where the rule can be inserted at the next iteration. In this particular 

implementation, F stands for moving forward, '/' and '\' for rotating a particular number of degrees, and '[' + ']' for branching (so

if we open a branch and close it we are back to the point where the branch started). So a string like A=FA/AF would draw a line

up, and a line sideways. When we iterate, we would get FFA/AF/FA/AFF. And so on, substituting the original rule for A everytime.

	The idea to move to 3D is not an original one, but not a lot of standard implementations exist for it. All that the system

has to have in order to be 3D is an additional pair of commands such as '+' and '-' for rotating in another axis in addition to '/' and '\'.

While 3d L-Systems partly resolve the problem, they require knowing the correct axioms/rules for the right effect. Thus

the designed is again faced with trial and error problem. 

	The solution in this case, much like in many other technological advances, is taken from observing nature. 
The technique suggested by the Wildwood paper is to take two L-systems (either designed or randomly generated) and 
to "breed" them to create a new L-System. This "breeding" would be a simulation of the natural evolution

process in nature. The "breeding" would occur by randomly selecting segments from either plant (meaning fragments of the strings

representing the plants) and also allowing for possibility of mutation (that is introduction of random strings into the offspring). 

	So the idea is to create an application that would allow random generation of samples and "breeding" to achieve

the desired plant without the complexity of 3D modeling or trial and error with L-System strings.

 Summary:
    - A need for a system for randomly generating and evolving (by fitness function or hand breeding)
       a set of L-systems-based plants
    - Would be useful in 3D model generation for trees by approximating realistic growth, but allowing
       a lot more control over the final model
 

	This approach is very similar to genetic algorithms and in fact it is the exact application of the genetic algorithm in

finding the 'right' plant. Genetic algorithms also generate a sample set of answers and select the 'best' by using a fitness function

that is applied to every answer. The 'best' answers are then cross-bred and a new set of samples is generated. This process

continues until a good enough answer is reached. This is exactly what virtual garden does, except the sample set consists of

different L-systems and the fitness function could be just selection by user (in addition to other automated fitness functions).

	Virtual Garden first of all implements a frame work to display 3D L-Systems from a given rule using Direct X. This 
is done by generating primitives for every command (so movement forward would be a cut-off cone and leafs are simple 
diamond shapes). Also a matrix stack structure is implemented for branches. 
	The primary feature of the application of course is the plant breeding part. Once a sample is generated using the Setup 
screen by either random generation or mutation of an original formula (see documentation below), the user is able to select

the next pair of plants to breed (or the fitness function selects automatically). The "breeding" is again accomplished by a technique 
similar to Wildwood. However, in order to increase visual appeal, a restraint is placed on what segments of each L-System

can be selected. Only entire branches are selected from each system. This still ensures proper cross-breeding but it also 

prevents the appearing of unlikely odd-looking systems (this has been determined by trial and error). 
	The user repeats this process until a desired plant is produced at which point the plant can be seen in different

number of iterations and it can be exported to 3D modeling software to be placed in a scene. 

Summary:

- Fast, easy, and 3D rather than previous related work dealing with 2D evolving plants
- User controlled or automatic breeding based on fitness function (height, width)
- More complex mutation / crossover functions to account for 3D grammar (branch based)
- Geared towards being a tool (color selectors, leaf shape selections, express interface)
- Preview of all generated samples at once
- Export to real-world apps
- Pretty realistic results 

General Plant Generation Setup
 

Going through the selection (manual or hand bred) as many times as needed
 

Changing the number of iterations on the final model and exporting it
 

Exported VRML model inside a browser