This guide has three parts:
BehaviorSpace is a software tool integrated with NetLogo that allows you to perform experiments with models.
BehaviorSpace runs a model many times, systematically varying the model's settings and recording the results of each model run. This process is sometimes called "parameter sweeping". It lets you explore the model's "space" of possible behaviors and determine which combinations of settings cause the behaviors of interest.
If your computer has multiple processor cores, then by default, model runs will happen in parallel, one per core.
The need for this type of experiment is revealed by the following observations. Models often have many settings, each of which can take a range of values. Together they form what in mathematics is called a parameter space for the model, whose dimensions are the number of settings, and in which every point is a particular combination of values. Running a model with different settings (and sometimes even the same ones) can lead to drastically different behavior in the system being modeled. So, how are you to know which particular configuration of values, or types of configurations, will yield the kind of behavior you are interested in? This amounts to the question of where in its huge, multi-dimension parameter space does your model perform best?
For example, suppose you want speedy synchronization from the agents in the Fireflies model. The model has four sliders -- number, cycle-length, flash-length and number-flashes -- that have approximately 2000, 100, 10 and 3 possible values, respectively. That means there are 2000 * 100 * 10 * 3 = 600,000 possible combinations of slider values! Trying combinations one at a time is hardly an efficient way to learn which one will evoke the speediest synchronization.
BehaviorSpace offers you a much better way to solve this problem. If you specify a subset of values from the ranges of each slider, it will run the model with each possible combination of those values and, during each model run, record the results. In doing so, it samples the model's parameter space -- not exhaustively, but enough so that you will be able to see relationships form between different sliders and the behavior of the system. After all the runs are over, a dataset is generated which you can open in a different tool, such as a spreadsheet, database, or scientific visualization application, and explore.
By enabling you to explore the entire "space" of behaviors a model can exhibit, BehaviorSpace can be a powerful assistant to the modeler.
To begin using BehaviorSpace, open your model, then choose the BehaviorSpace item on NetLogo's Tools menu.
The dialog that opens lets you create, edit, duplicate, delete, and run experiment setups. Experiments are listed by name and how by model runs the experiment will consist of.
Experiment setups are considered part of a NetLogo model and are saved as part of the model.
To create a new experiment setup, press the "New" button.
In the new dialog that appears, you can specify the following information. Note that you don't always need to specify everything; some parts can be left blank, or left with their default values, depending on your needs.
Experiment name: If you have multiple experiments, giving them different names will help you keep them straight.
Vary variables as follows: This is where you specify which settings you want varied, and what values you want them to take. Variables can include sliders, switches, choosers, and any global variables in your model.
Variables can also include
random-seed. These are
not, strictly speaking, variables, but BehaviorSpace lets you vary
them as if they were. Varying the world dimensions lets you explore
the effect of world size upon your model. Since setting
world-height does not
necessarily define the bounds of the world how they are varied
depends on the location of the origin. If the origin is centered,
BehaviorSpace will keep it centered so the values
world-height must be odd.
If one of the bounds is at zero that bound will be kept at zero and
the other bound will move, for example if you start with a world with
min-pxcor = 0
max-pxcor = 10 and you vary
world-width like this:
["world-width" [11 1 14]]
min-pxcor will stay
at zero and
max-pxcor will set to 11,
12, and 13 for each of the runs. If neither of these conditions are
true, the origin is not centered, nor at the edge of the world you
world-width directly but
you should vary
random-seed lets you
repeat runs by using a known seed for the NetLogo random number
generator. Note that you're also free to use the
random-seed command in
your experiment's setup commands. For more information on random
seeds, see the Random Numbers
section of the Programmer's Guide.
You may specify values either by listing the values you want used, or
by specifying that you want to try every value within a given range.
For example, to give a slider named
number every value from
100 to 1000 in increments of 50, you would enter:
["number" [100 50 1000]]
Or, to give it only the values of 100, 200, 400, and 800, you would enter:
["number" 100 200 400 800]
Be careful with the brackets here. Note that there are fewer square brackets in the second example. Including or not including this extra set of brackets is how you tell BehaviorSpace whether you are listing individual values, or specifying a range.
Also note that the double quotes around the variable names are required.
You can vary as many settings as you want, including just one, or none at all. Any settings that you do not vary will retain their current values. Not varying any settings is useful if you just want to do many runs with the current settings.
What order you list the variables in determines what order the runs will be done in. All values for a later variable will be tried before moving to the next value for an earlier variable. So for example if you vary both x and y from 1 to 3, and x is listed first, then the order of model runs will be: x=1 y=1, x=1 y=2, x=1 y=3, x=2 y=1, and so on.
Repetitions: Sometimes the behavior of a model can vary a lot from run to run even if the settings don't change, if the model uses random numbers. If you want to run the model more than once at each combination of settings, enter a higher number.
Measure runs using these reporters: This is where you specify what data you want to collect from each run. For example, if you wanted to record how the population of turtles rose and fell during each run, you would enter:
You can enter one reporter, or several, or none at all. If you enter several, each reporter must be on a line by itself, for example:
If you don't enter any reporters, the runs will still take place.
This is useful if you want to record the results yourself your own
way, such as with the
Measure runs at every step: Normally NetLogo will measure model runs at every step, using the reporters you entered in the previous box. If you're doing very long model runs, you might not want all that data. Uncheck this box if you only want to measure each run after it ends.
Setup commands: These commands will be used to begin each
model run. Typically, you will enter the name of a procedure that
sets up the model, typically
setup. But it is also possible
to include other commands as well.
Go commands: These commands will be run over and over again to
advance to the model to the next "step". Typically, this
will be the name of a procedure, such as
go, but you may
include any commands you like.
Stop condition: This lets you do model runs of varying length, ending each run when a certain condition becomes true. For example, suppose you wanted each run to last until there were no more turtles. Then you would enter:
not any? turtles
If you want the length of runs to all be of a fixed length, just leave this blank.
The run may also stop because the go commands use the
stop command, in the same way
stop can be used to
stop a forever button. The
stop command may be used directly
in the go commands, or in a procedure called directly by the go
commands. (The intent is that the same
go procedure should
work both in a button and in a BehaviorSpace experiment.) Note that
the step in which
used is considered to have been aborted, so no results will be
recorded for that step. Therefore, the stopping test should be at the
beginning of the go commands or procedure, not at the end.
Final commands: These are any extra commands that you want run
once, when the run ends. Usually this is left blank, but you might
use it to call the
export-world command or
record the results of the run in some other way.
Time limit: This lets you set a fixed maximum length for each run. If you don't want to set any maximum, but want the length of the runs to be controlled by the stop condition instead, enter 0.
Currently there are only two,
The run number reported by the former primitive matches the run number
used in the results files generated by BehaviorSpace. The experiment
name reported by the latter matches the name with which the experiment
was set up.
When you're done setting up your experiment, press the "OK" button, followed by the "Run" button. A dialog titled "Run options" will appear.
The run options dialog lets you select the formats you would like the data from your experiment saved in. Data is collected for each run or step, according to the setting of Measure runs at every step option. In either case, the initial state of the system is recorded, after the setup commands run but before the go commands run for the first time.
Table format lists each interval in a row, with each metric in a separate column. Table data is written to the output file as each run completes. Table format is suitable for automated processing of the data, such as importing into a database or a statistics package.
Spreadsheet format calculates the min, mean, max, and final values for each metric, and then lists each interval in a row, with each metric in a separate column. Spreadsheet data is more human-readable than Table data, especially if imported into a spreadsheet application.
(Note however that spreadsheet data is not written to the results file until the experiment finishes. Since spreadsheet data is stored in memory until the experiment is done, very large experiments could run out of memory. So you should disable spreadsheet output unless you really want it. If you do want spreadsheet output, note that if anything interrupts the experiment, such as a runtime error, running out of memory, or a crash or power outage, no spreadsheet results will be written. For long experiments, you may want to also enable table format as a precaution so that if something happens and you get no spreadsheet output you'll at least get partial table output.)
After selecting your output formats, BehaviorSpace will prompt you for the name of a file to save the results to. The default name ends in ".csv". You can change it to any name you want, but don't leave off the ".csv" part; that indicates the file is a Comma Separated Values (CSV) file. This is a plain-text data format that is readable by any text editor as well as by most popular spreadsheet and database programs.
The run options dialog also lets you select whether you want multiple model runs to happen in parallel, and if so, how many are allowed to be simultaneously active. This number will default to the number of processor cores in your computer.
There are a few cautions associated with parallel runs.
First, if multiple runs are active, only one of them will be in the "foreground" and cause the view and plots to update. The other runs will happen invisibly in the background.
Second, invisible background runs can't use primitives that only work in the GUI. For example, a background run can't make a movie.
Third, since parallel runs progress independently of each other, table format output may contain interleaved, out-of-order results. When you analyze your table data, you may wish to sort it by run number first. (Spreadsheet format output is not affected by this issue, since it is not written until the experiment completes or is aborted.)
Fourth, using all available processor cores may make your computer slow to use for other tasks while the experiment is running.
Fifth, doing runs in parallel will multiply the experiment's memory requirements accordingly. You may need to increase NetLogo's memory ceiling (see this FAQ entry).
After you complete the run options dialog, another dialog will appear, titled "Running Experiment". In this dialog, you'll see a progress report of how many runs have been completed so far and how much time has passed. If you entered any reporters for measuring the runs, and if you left the "Measure runs at every step" box checked, then you'll see a plot of how they vary over the course of each run.
You can also watch the runs in the main NetLogo window. (If the "Running Experiment" dialog is in the way, just move it to a different place on the screen.) The view and plots will update as the model runs. If you don't need to see them update, then use the checkboxes in the "Running Experiment" dialog to turn the updating off. This will make the experiment go faster.
If you want to stop your experiment before it's finished, press the "Abort" button. Any results generated so far will still be saved.
When all the runs have finished, the experiment is complete.
It is possible to run BehaviorSpace experiments "headless", that is, from the command line, without any graphical user interface (GUI). This is useful for automating runs on a single machine or a cluster of machines.
No Java programming is required. Experiment setups can be created in the GUI and then run later from the command line, or, if you prefer, you can create or edit experiment setups directly using XML.
Run Java with the org.nlogo.headless.Main class. The Main.main() method supports these arguments:
--model <path>: pathname of model to open (required)
--setup-file <path>: read experiment setups from this file instead of the model file
--experiment <name>: name of experiment to run
--table <path>: pathname to send table output to (or
-for standard output)
--spreadsheet <path>: pathname to send table output to (or
-for standard output)
--threads <number>: use this many threads to do model runs in parallel, or 1 to disable parallel runs. defaults to one thread per processor.
--min-pxcor <number>: override world size setting in model file
--max-pxcor <number>: override world size setting in model file
--min-pycor <number>: override world size setting in model file
--max-pycor <number>: override world size setting in model file
--model is required. If you don't specify
--experiment, you must specify
default no results are generated, so you'll usually want to
--spreadsheet, or both.
If you specify any of the world dimensions, you must specify all
It is easiest if you create your experiment setup ahead of time in the GUI, so it is saved as part of the model. To run an experiment setup saved in a model, here is an example command line:
java -Xmx1024m -Dfile.encoding=UTF-8 -cp NetLogo.jar \ org.nlogo.headless.Main \ --model Fire.nlogo \ --experiment experiment1 \ --table -
(For this to work,
NetLogo.jar must be present along with
lib subdirectory containing necessary libraries. Both
lib are included with NetLogo.)
After the named experiment has run, the results are sent to standard output in table format, as CSV. ("-" is how you specify standard output instead of output to a file.)
When running the headless.Main class as an application, it forces the
java.awt.headless to be true. This tells
Java to run in headless mode, allowing NetLogo to run on machines
when a graphical display is not available.
Note the use of
-Xmx to specify a maximum heap size of one
gigabyte. If you don't specify a maximum heap size, you will get
your VM's default size, which may be unusably small. (One
gigabyte is an arbitrary size which should be more than large enough
for most models; you can specify a different limit if you want.)
Note the use of
-Dfile.encoding=UTF-8. This forces all file
I/O to use UTF-8 encoding. Doing so ensures that NetLogo can load all
models consistently, and that file-* primitives work consistently on
all platforms, including models containing Unicode characters.
--model argument is used to specify the model
file you want to open.
--experiment argument is used to specify the name of the
experiment you want to run. (At the time you create an experiment
setup in the GUI, you assign it a name.)
Here's another example that shows some additional, optional arguments:
java -Xmx1024m -Dfile.encoding=UTF-8 -cp NetLogo.jar \ org.nlogo.headless.Main \ --model Fire.nlogo \ --experiment experiment2 \ --max-pxcor 100 \ --min-pxcor -100 \ --max-pycor 100 \ --min-pycor -100
Note the use of the optional
--max-pycor, etc. arguments to specify a different world
size than that saved in the model. (It's also possible for the
experiment setup to specify values for the world dimensions; if they
are specified by the experiment setup, then there is no need to
specify them on the command line.)
Since neither --table nor --spreadsheet is specified, no results will be generated. This is useful if the experiment setup generates all the output you need by some other means, such as exporting world files or writing to a text file.
Yet another example:
java -Xmx1024m -Dfile.encoding=UTF-8 -cp NetLogo.jar \ org.nlogo.headless.Main \ --model Fire.nlogo \ --experiment experiment2 \ --table table-output.csv \ --spreadsheet spreadsheet-output.csv
--table <filename> argument specifies
that output should be generated in a table format and written to the
given file as CSV data. If
- is specified as the filename,
than the output is sent to the standard system output stream. Table
data is written as it is generated, with each complete run.
--spreadsheet <filename> argument
specified that spreadsheet output should be generated and written to
the given file as CSV data. If
- is specified as the
filename, than the output is sent to the standard system output
stream. Spreadsheet data is not written out until all runs in the
experiment are finished.
Note that it is legal to specify both
--spreadsheet, and if you do, both kinds of output file will
Here is one final example that shows how to run an experiment setup which is stored in a separate XML file, instead of in the model file:
java -Xmx1024m -Dfile.encoding=UTF-8 -cp NetLogo.jar \ org.nlogo.headless.Main \ --model Fire.nlogo \ --setup-file fire-setups.xml \ --experiment experiment3
If the XML file contains more than one experiment setup, it is
necessary to use the
--experiment argument to specify the
name of the setup to use.
In order to run any of these experiments in 3D add
-Dorg.nlogo.is3d=true to any of these startup commands, for
java -Dorg.nlogo.is3d=true -Xmx1024m -Dfile.encoding=UTF-8 -cp NetLogo.jar \ org.nlogo.headless.Main \ --model Fire3D.nlogo \ --experiment experiment1 \ --table -
Note that you should supply a 3D model and there are also 3D
--max-pzcor <number> and
The next section has information on how to create standalone experiment setup files using XML.
We don't yet have detailed documentation on authoring experiment setups in XML, but if you already have some familiarity with XML, then the following pointers may be enough to get you started.
The structure of BehaviorSpace experiment setups in XML is determined
by a Document Type Definition (DTD) file. The DTD is stored in
system/behaviorspace.dtd. (JAR files are
also zip files, so you can extract the DTD from the JAR using
Java's "jar" utility or with any program that
understands zip format.)
The easiest way to learn what setups look like in XML, though, is to
author a few of them in BehaviorSpace's GUI, save the model, and
then examine the resulting .nlogo file in a text editor. The
experiment setups are stored towards the end of the .nlogo file, in a
section that begins and ends with a
<experiments> <experiment name="experiment" repetitions="10" runMetricsEveryStep="true"> <setup>setup</setup> <go>go</go> <exitCondition>not any? fires</exitCondition> <metric>burned-trees</metric> <enumeratedValueSet variable="density"> <value value="40"/> <value value="0.1"/> <value value="70"/> </enumeratedValueSet> </experiment> </experiments>
In this example, only one experiment setup is given, but you can put
as many as you want between the beginning and ending
Between looking at the DTD, and looking at examples you create in the GUI, it will hopefully be apparent how to use the tags to specify different kind of experiments. The DTD specifies which tags are required and which are optional, which may be repeated and which may not, and so forth.
When XML for experiment setups is included in a model file, it does not begin with any XML headers, because not the whole file is XML, only part of it. If you keep experiment setups in their own file, separate from the model file, then the extension on the file should be .xml not .nlogo, and you'll need to begin the file with proper XML headers, as follows:
<?xml version="1.0" encoding="UTF-8"?> <!DOCTYPE experiments SYSTEM "behaviorspace.dtd">
The second line must be included exactly as shown. In the first line,
you may specify a different encoding than
UTF-8, such as
If BehaviorSpace is not sufficient for your needs, a possible alternative is to use our Controlling API, which lets you write Java code that controls NetLogo. The API lets you run BehaviorSpace experiments from Java code, or, you can write custom code that controls NetLogo more directly to do BehaviorSpace-like things. See the Controlling section of the User Manual for further details on both possibilities.