Skip to content
/ pasta Public

Agent-based simulation with evolutionary strategies, individual and social learning, simple model of prediction error minimization

License

Notifications You must be signed in to change notification settings

mars0i/pasta

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

pasta

Marshall Abrams, http://members.logical.net/~marshall

Sections: Overview, How to run it, How it works, Parameters & command line options, Things to try, More info, Full installation, License

Overview:

This is a model of evolutionary competition between different strategies for dealing with variable environments. It's an agent-based simulation1. There are three kinds of agents (models of organisms) that can be deployed:

k-snipes: Agents that incorporate a very simple model of individual learning by prediction error minimization2 (PEM) in some agents. These agents implement one kind of "K-strategy" which, roughly, prioritizes survival over reproduction. (I do mean it's a model of PEM; what these agents do, internally, is much, much simpler than what's usually meant by PEM.)

r-snipes: Agents that don't learn, but produce different types of offspring that are well suited or poorly suited for survival in different environments. These agents use "bet hedging" to implement one kind of "r-strategy" which, roughly, prioritizes reproduction over survival, "betting" on the possibilities of different environmental states.

s-snipes: Agents that engage in a simple form of social learning by copying from nearby agents.

The point is to compare evolutionary strategies and see which is selected for under different parameter combinations.

How to run it:

Download pasta.jar to your computer. If your computer is configured appropriately, you should be able to double-click or click on this file to run pasta. If not, and you're comfortable with a terminal window (OSX terminal, Windows cmd.exe, any shell window in Linux), you can run pasta using this command:

java -jar pasta.jar

If you're not in the directory where you've put pasta.jar, you might have to add a path to the filename. You can use command line options to run pasta in no-GUI mode by executing the line above and adding --help.

For other ways to run pasta, see "Full installation" below.

How it works:

(This doesn't document features of the GUI that are probably easy to figure out with a bit of guessing and trial and error.)

The two environments: Snipes eat mushrooms.3 In the east (usually left) environment, small mushrooms are nutritious (gray-brown-green) and large mushrooms are slightly poisonous (greenish yellow). In the west (usually right) environment, it's the large mushrooms that are nutritious (dark gray), and the small mushrooms that are poisonous (light gray). The general rule is that darker mushrooms are nutritious and lighter mushrooms are poisonous.

Basic snipe behavior: Snipes move randomly within an environment. Note that the orientation of snipes' icons does not indicate direction of movement (see below). Snipes gain energy from eating nutritious mushrooms, and lose energy from eating poisonous mushrooms or from producing an offspring. Movement and internal processes have no cost. Snipes that accumulate sufficient energy will give birth to a single offspring, losing energy as a result. A snipe can have multiple offspring only by repeatedly acquiring enough energy to give birth. At birth, each newborn snipe is placed at a random location in a randomly chosen environment, as if parents had temporarily migrated to a new location to leave an egg to hatch.

k-snipes (red circles with pointers) initially eat mushrooms randomly, i.e. with no preference for large or small mushrooms, but learn to eat mushrooms whose size signal (which is normally distributed) indicates that they are probably nutritious. The direction of a k-snipe's pointer—how far up or down it is pointing—indicates the degree of the snipe's preference for large or small mushrooms. For details, see doc/kSnipePerception.pdf.

r-snipes (blue triangles) never learn. They produce offspring that exhibit developmental differences: Roughly half of any snipe's offspring (upward-pointing triangles) always prefer large mushrooms; the others (downward-pointing triangles) always prefer small mushrooms. Those suited to the environment in which they live tend to survive and reproduce, and those unsuited to their environment generally die before reproduction. This is a kind of {\em bet hedging\/} strategy: Rather than "betting" on a condition (that large mushrooms are nutritious, for example), r-snipes hedge their bets by placing bets (in the form of offspring) on both environmental patterns.

s-snipes (purple wing shapes) use a social learning or cultural transmission strategy known as success bias. A newborn s-snipe examines nearby snipes and copies the current mushroom size preference of whichever nearby snipe has the most energy. If there are no snipes that are sufficiently near (see the parameter list below), the s-snipe tries again on the next time step. Once an s-snipe adopts a preference, the preference never changes. The direction in which an s-snipe points—how far up or down it is tilted—indicates the degree of the snipe's preference for large or small mushrooms.

Colors and energy: Snipes' energy levels are reflected in their brightness, with greater brightness indicating more energy. (This effect can be subtle.) The two mushroom colors in each environment indicate nutritiousness/poisonousness: Darker colors indicate nutritiousness—i.e. these mushrooms are energetically favorable to snipes—while lighter colors indicate poisonousness—energetic unfavorability. (The fact that the East and West mushrooms have different hues has no functional meaning; it could be considered a difference between local mushroom species.)

More on energy: In this simple model the only things that can reduce the energy of a snipe are (a) eating poisonous mushrooms, or (b) producing offspring. There is no cost to movement or simple persistence. You can assume that snipes have another source of nutrition that maintains them, but that is not represented in the model. In theory a snipe could live forever without eating any nutritious mushrooms, as long as it never ate poisonous mushrooms. (Such a snipe would never give birth more than once, either; if it had enough energy to give birth it would, but its energy would be reduced as a result and it would never acquire any additional energy since it never ate nutrious mushrooms.) Assume that this other source of nutritiou would not on its own allow a snipe to gain sufficient energy for reproduction. Also, although in nature, mechanisms that allow learning can be costly in energy use or extra developmental time, k-snipe's individual learning and s-snipe's social learning has no energetic or time cost in the model. k-snipes do suffer a cost relative to r-snipes, because during a k-snipe's learning process, it often ends up eating a number of poisonous mushrooms. This slows down its acquisition of energy for birth, and sometimes leads to a sustained loss of energy during its initial learning period. (A k-snipe never stops learning, but after a while it will mostly choose nutritious mushrooms, and its preference for whatever mushroom size is nutritious in its environment is merely reinforced.) Whether an s-snipe's social learning ends up being costly or profitable depends on its neighbors early in its life.

Monitoring individual snipes: If you pause a run, you can double-click on a snipe to monitor its internal state and watch it move. It will be circled so that you can keep track of it. You can use the Detach button to monitor multiple snipes at the same time. If you double-click on the little magnifying glass next to energy, for example, and choose chart, you'll open a plot of that snipe's energy level over time.

Overlapping environment display: It's possible to display both environments overlapping in the same window, using the "overlapping subenvs" item on the Displays tab. Then you can think of snipes that only eat mushrooms of a given color as having a random developmental restriction to eating those mushrooms.

Command line: You can also run pasta from a command line (see online documents or README.md in the full distribution). In that case, running it with "--help" will show options for running pasta.

Parameters & command line options:

Parameters in the GUI:

The following parameters for model runs can be set in the Model tab:

NumKSnipes: Size of k-snipe subpopulation
NumRSnipes: Size of r-snipe subpopulation
NumSSnipes: Size of s-snipe subpopulation
MushProb: Average frequency of mushrooms.
MushHighSize: Size of large mushrooms (mean of light distribution)
MushLowSize: Size of small mushrooms (mean of light distribution)
MushSd: Standard deviation of mushroom light distribution
MushPosNutrition: Energy from eating a nutritious mushroom
MushNegNutrition: Energy from eating a poisonous mushroom
InitialEnergy: Initial energy for each snipe
BirthThreshold: Energy level at which birth takes place
KPrefNoiseSd: Standard deviation of internal noise in k-snipe preference determination.
BirthCost: Energetic cost of giving birth to one offspring
MaxEnergy: Max energy that a snipe can have.
Lifespan: Each snipe dies after this many timesteps.
CarryingProportion: Snipes are randomly culled when number exceed this times # of cells in a subenv (east or west).
NeighborRadius: s-snipe neighbors (for copying) are no more than this distance away.
EnvWidth: Width of env. Must be an even number.
EnvHeight: Height of env. Must be an even number.
ExtremePref: Absolute value of r-snipe preferences.
ReportEvery: Report basic stats every i ticks after the first one (0 = never); format depends on -w.
KMaxPopSizes: Comma-separated times and target subpop sizes to cull k-snipes to, e.g. "time,size,time,size"
RMaxPopSizes: Comma-separated times and target subpop sizes to cull r-snipes to, e.g. "time,size,time,size"
SMaxPopSizes: Comma-separated times and target subpop sizes to cull s-snipes to, e.g. "time,size,time,size"
KMinPopSizes: Comma-separated times and target subpop sizes to increase k-snipes to, e.g. "time,size,time,size"
RMinPopSizes: Comma-separated times and target subpop sizes to increase r-snipes to, e.g. "time,size,time,size"
SMinPopSizes: Comma-separated times and target subpop sizes to increase s-snipes to, e.g. "time,size,time,size"

Some data is reported in other variables displayed in the Model tab:

PopsSize Number of snipes in the population (in both environments).
KSnipeFreq Relative frequency of k-snipes.
RSnipeFreq Relative frequency of r-snipes.
SSnipeFreq Relative frequency of s-snipes.

If you click on the little magnifying glass next to these elements, there are additional options that MASON provides for observing their values.

Command line options:

If you run pasta from the command line `-?` or `--help` will show options for setting the parameters listed above, as well as a few other options concerning whether to display the GUI concerning writing summary data to a file. It's also useful to run pasta with `-help` (one dash) to see additional command line options provided by the MASON library that pasta uses. One particularly useful MASON options is `-for `, which specifies the number of time steps to run. Without `-for` (or `-until`), pasta will run forever, unless you suspend it or kill it.

Here is more information about making pasta generate data:

If you add -i <integer> or --report-every <integer> to the command line, pasta will output summary statistics on different classes of snipes every <integer> timesteps. If you also add -w or --write-csv, pasta will write these statistics to a file. It will also write a separate file containing the parameters for this run. If you add -F <name> or --csv-basename <name> as well, pasta will use <name> as the beginning of the filenames. Stats are also written on the last timestep, whether it's a multiple of the value for -i or --report-every, as long as that value is greater than zero. This means that if you only want stats on the last time step, simply give -i or --report-every a value greater than the one given to -for.

If the data is written to the file, the resulting csv file (which can be pulled into Excel, for example) will consist of rows of data separated by commas. There will be a header row in a separate file with "header" in its name which you can concatenate onto the data file. The columns one for the run id (which is also the random seed), so that you can append multiple runs to the same file; the time step at which the data was collected; the snipe class—i.e. whether it is a k-snipe, r-snipe, or s-snipe; the sub-environment (east or west) of the snipes that are summarized; whether the snipes summarized have a negative, positive, or neutral size preference for mushrooms; the number of mushrooms in the condition specified by the previous columns; their average energy; average mushroom preference value; and average age. If you use the same basename as an existing data file, or if you specify a basename with multiple runs using MASON's -repeat, the data will be appended to the file, and a single parameters file will be written; otherwise, every run's data will have unique filenames. (Note: A snipe's mushroom preference value is a real number that is zero to indicate that the snipe has no preference for large vs. small mushrooms, positive to indicate the degree of preference for large mushrooms, and negative to indicate the degree of preference for small mushrooms. See doc/kSnipePerception.pdf for details.)

If the data isn't written to a file, it will be sent to standard output. The format for the data sent to is different though. It's not it's not ideally user-friendly, and I need to document it better. (The format is defined somewhere in the guts of stats.clj.) If you prefer the csv format but want to see data immediately as it's generated, you can write it to a file and then keep looking at that file; on Unix systems, the tail command is useful for this purpose.

(If you want more fine-grained data, you could modify the source code for pasta—I would be happy to help you if I have time—but I would suggest simply using the recorded seed to run the same simulation in the GUI. You can use the GUI to inspect individual snipes that way, as indicated above.)

Things to try:

Set the number of s-snipes (NumSSnipes) or the number of r-snipes (NumRSnipes) to zero, so that there is only competition between k-snipes and one of the other varieties. Notice that with the default parameters, k-snipes sometimes increase in frequency early on, but in the end the r-snipes or s-snipes always increase in frequency at the expense of k-snipes. Why? (Consider monitoring the energy levels of a few snipes over time.)

Experiment with parameters to see if you can cause k-snipes to win the evolutionary race. For example, before the start of a k-snipe vs. r-snipe run, reduce MushNegNutrition. Does this affects k-snipes' relative success? Why?

More info:

The name of this project comes from the r-snipes' strategy of producing lots of offspring, many of whom will be maladapted to their environment and die. Kind of like testing pasta by throwing it against the wall to see what sticks.

As noted above, k-snipes implement a certain kind of evolutionary K strategy, while r-snipes implement a certain kind of evolutionary r strategy. (These terms are not always defined in precisely the same ways.)

Note that in this model, the K vs r strategies have nothing to do with K vs r selection in the sense of responding to having or not having bounded resources or population size. The simulation produces qualitatively similar results whether the population is growing or has reached its maximum size, for example.

My starting points on PEM are Rafal Bogacz, "A tutorial on the free-energy framework for modeling perception and learning", Journal of Mathematical Psychology 76 Part B, Feb. 2017, pp. 198-211, http://dx.doi.org/10.1016/j.jmp.2015.11.003, and Harriet Feldman and Karl Friston, "Attention, Uncertainty, and Free-Energy", Frontiers in Human Neuroscience 4, 2010, http://dx.doi.org/10.3389/fnhum.2010.00215. However, the model implemented here in the k-snipe agents is very simple and very different. Andy Clark's Surfing Uncertainty (Oxford 2015) and Jakob Hohwy's The Predictive Mind (Oxford 2014) contain non-technical discussions of PEM.

Full installation

This program is written in the Clojure language using MASON. MASON is a Java library for agent-based modeling.

Installation:

First clone this git repo. (If you're not sure how to do this, you should be able to find beginner info about git and github on the web or in books written about them.)

You need install Leiningen (http://leiningen.org). Then change to the root directory of the repo, and run 'lein deps'. This will install the right version of Clojure, if necessary, along with a number of Clojure libraries and such that pasta needs.

You'll also need to download the MASON jar file mason.19.jar (or a later version, probably), and MASON libraries.tar.gz or libraries.zip file. Move the MASON jar file into the lib directory under this project's directory. Unpack the contents of the libraries file into this directory as well. You may want to get the MASON-associated 3D libs, too, in order to get rid of a warning message during startup. Place these jars in the same places as the others.

Ways to run pasta using the full installation:

There is another way to run it described above in "How to run it".

These instructions were tested on MacOS, and they should work on any full-fledged unix system such as Linux, MacOS, etc. The same instructions might work on Windows, though you might need to substitute "" for "/".

(1) From the repository directory execute src/scripts/gui or src/scripts/pasta will start the GUI version of pasta.

(2) Running src/scripts/nogui, or running src/scripts/pasta with no-GUI command line options will start the command line version (unless -g is specified). You may want to run it with -? as argument first to see the possible command line options.

(3) Running src/scripts/repl will start the Clojure REPL. This is the same as running lein repl, but displays some helpful suggestions. e.g. it suggests that you can start the GUI using some variation on the following Clojure commands:

(use 'pasta.UI) (def cfg (repl-gui))

If you run these, the GUI should start up, and you can run the program from there. However, you can also use the REPL to examine the state of the program:

(def cfg-data$ (.simConfigData cfg))

This defines cfg-data$ as a Clojure atom containing structure (essentially a Clojure map) with various sorts of parameters and runtime state. (I follow a non-standard convention of naming variables containing atoms with a dollar sign character as suffix.) For example, all currently living snipes are listed in map called :snipes, keyed by id, in the :popenv structure in the structure to which cfg-data$ refers. For example, since ids are assigned sequentially, the largest id is the count of all snipes that have lived:

(apply max (keys (:snipes (:popenv @cfg-data$))))

Warning: Unfortunately, snipes contain a reference to cfg-data$ itself, and by default the REPL will try to list the contents of atoms, so if you allow any snipe to print to the terminal, you'll set off an infinite loop that will result in a stack overflow. Sorry about that!

License:

This software is copyright 2016, 2017, 2018, 2019 by Marshall Abrams, and is distributed under the Gnu General Public License version 3.0 as specified in the file LICENSE, except where noted, or where code has been included that was released under a different license.

footnotes:

1 "Agent-based" here refers to the loose class of simulations in which outcomes of interest come from interactions between many semi-independent entities—agents—which often model behaviors or interactions between people, organisms, companies, etc. "Agent-based" does not refer to various ways of dealing with concurrency, as for example with Clojure's agent data structure.

2 Prediction energy minimization: AKA free-energy minimization, predictive processing, predictive coding. cf. integral control, mean-field approximation, variational methods.

3Yes—this pasta has mushrooms in it. Yum!

About

Agent-based simulation with evolutionary strategies, individual and social learning, simple model of prediction error minimization

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published