This is a model of the <a href="https://en.wikipedia.org/wiki/Iowa_caucuses" target="_blank" rel="noopener">Iowa caucuses</a>, a political event that combines voting and debate/discussion. In this model, voters and candidates are represented on a political compass, with each point in space representing a unique ideological viewpoint, all drawn from the same distribution function. Voters are also connected to each other, representing social relationships. They are connected to their closest neighbors ideologically, with random connections drawn to voters on other parts of the political spectrum.
The caucus proceeds like this: In the first round, each voter casts their vote for the candidate whose ideological location is &#8216;closest&#8217; to theirs. Then, in following rounds, candidates receiving less than 15% of the total vote are marked not &#8216;viable,&#8217; and voters who chose a nonviable candidate vote again by randomly picking one of the candidates chosen by their connections (including the same candidate). This continues until all candidates receiving votes have over 15% of the total.
By situating the landscape of voters on a social network, it is possible for candidates receiving less than 15% of the vote in the first round to become viable candidates through their voters influencing others in subsequent rounds.

Iowa Caucuses

Starting with no common language, a group of agents learn to communicate with each other through a guessing game. Each turn, an agent points to a unique object in a group of three objects. A second agent guesses the &#8216;meaning&#8217; behind the first agent&#8217;s act of pointing — a distinguishing feature, like &#8220;blue,&#8221; &#8220;circle,&#8221; or &#8220;1st&#8221; — and produces a word for that meaning. If it matches the word the first agent had in mind, the association between the word and the meaning are strengthened (however, in ambiguous contexts, it&#8217;s possible for the second agent to match the first agent&#8217;s word but for a different meaning). At the start, since the agents have no vocabulary, they generate new words and use them randomly, but over time, a shared language emerges that allows them to successfully communicate about the features of objects.
Despite the simplicity of this model, a number of complex linguistic phenomena emerge, including synonyms (multiple words with the same meaning), differing usage/&#8217;slang&#8217; among sub-populations, and context-specificity (ex. a word that is associated with &#8220;blue square,&#8221; but not squares in general). This model is based on Luc Steels and Frédéric Kramer&#8217;s <a href="https://ai.vub.ac.be/sites/default/files/steels-ground1999.pdf">Bootstrapping Grounded Word Semantics</a> (1999).

Language Evolution

This model, based on observations of fireflies in Thailand, shows how a system of distributed entities with periodic behavior can eventually become nearly perfectly in sync. Each &#8216;firefly&#8217; agent has an internal clock, and when it flashes, nearby fireflies accelerate their clocks forward slightly. As the fireflies randomly navigate the environment, they encounter more and more flashing fireflies and grow more in sync with them. After a relatively short amount of time, all the fireflies are flashing in unison, with no top-down or centralized behavior organizing them. Like in the <a href="https://www.flocc.network/model/flocking">flocking model</a>, this behavior emerges out of the interactions of many individual agents.
The implementation of this model is indebted to Nicky Case&#8217;s interactive explorable <a href="https://ncase.me/fireflies/">&#8220;Fireflies&#8221;</a>, and the phenomenon is described in the opening chapter of Steven Strogatz&#8217;s book <a href="https://www.stevenstrogatz.com/books/sync-the-emerging-science-of-spontaneous-order">Sync: How Order Emerges from Chaos in the Universe, Nature, and Daily Life</a>.

Fireflies Syncing

The <a href="https://en.wikipedia.org/wiki/Lorenz_system">Lorenz system</a> is a set of three interrelated differential equations that, with certain starting parameters, can result in locally chaotic but globally predictable values. In this model, five agents start at random positions (between 0 and 1 for x, y, and z values) and follow the update rules of the Lorenz system. Although their paths are initially similar, they soon diverge and one agent&#8217;s position is totally unrelated to the next. However, the five paths trace the form of a Lorenz attractor — a set of solutions to the system within which any given agent&#8217;s position will lie. Since there are no feedback loops between agents interacting with one another, this is an example of a complex physical system rather than a complex adaptive system.
The visualization uses Three.js for rendering.

Lorenz Attractor

This model, outlined in Miller &amp; Page&#8217;s <a href="https://press.princeton.edu/books/paperback/9780691127026/complex-adaptive-systems">Complex Adaptive Systems</a>, represents an organization that makes &#8220;decisions&#8221; — binary outputs of either &#8220;0&#8221; or &#8220;1&#8221; — based on four-bit binary input &#8220;problems&#8221; (for example, &#8220;0110&#8221; or &#8220;1101&#8221;). The organization is structured with three agents, who each make decisions based on two-bit binary inputs (&#8220;00,&#8221; &#8220;01,&#8221; &#8220;10,&#8221; or &#8220;11&#8221;). They are all initialized with a &#8220;decision table&#8221; that determines what output they produce for each input. Two of the agents take input directly from the four-bit problem (the first two and the last two digits, respectively), and the third agent takes input from these agents&#8217; outputs, outputting the organization&#8217;s decision.
Since each problem has a &#8220;correct&#8221; solution, and there are 16 unique problem-solution pairs, it&#8217;s unlikely that the organization will correctly decide every problem. The organization&#8217;s baseline (percent of problems it will correctly decide) is displayed at the top, and the line chart updates to show a running tally of the mean percent of problems decided correctly.

Organizational Decision Making

This simple model reflects a phenomenon of two-party politics where both candidates shift their positions to garner more votes. Voters are <a href="https://en.wikipedia.org/wiki/Left%E2%80%93right_political_spectrum">represented along a single axis</a>, and two candidates are initially placed at random positions in the left half and the right half (blue and red, respectively, as in U.S. politics). A voter casts their ballot for the candidate whose position is closer to theirs.
After each round of voting, the candidate with fewer votes shifts their position (there is no incentive for an incumbent to adopt a new position). With LEARNING_RATE set to 1, they will inevitably move toward the center, with the opponent doing the same if and when they are the minority candidate. With a lower LEARNING_RATE, the process still moves toward the center, but resembles a random walk, where candidates try out new positions that may be less optimal and find equilibrium more slowly.

Race to the Center

In the Predator-Prey model, there are two agent types: Sheep and wolves. Sheep eat grass, and their population is limited by both their food source — above a certain carrying capacity there aren&#8217;t enough resources to go around — and by the wolves who prey on nearby sheep. The nonlinear dynamics of this model are extremely sensitive to the starting conditions and parameters of reproductive and metabolic rates.

Predator-Prey

In the minority game (based on the NetLogo model of the same name), agents make a guess of either 0 or 1 with each tick of the simulation. After each tick, the goal for each agent is to be in the minority (guessing the same as less than half of all other agents).
Agents are initialized with a set of strategies — binary strings, i.e. 010111011000&#8230; — and the history of recent winning (minority) guesses are stored as another binary string. A strategy &#8216;chooses&#8217; the agent&#8217;s guess by converting the history to a decimal number (i.e. 1101 → 13) and selecting that digit from the strategy (which is always either 0 or 1). After each tick, an agent updates the scores of all its strategies and selects the one with the highest score, using it for subsequent guesses. Although this behavior is somewhat opaque, it does result in unpredictable, chaotic behavior as agents attempt to guess at and outperform the market.

Minority Game

The Game of Life shows how remarkable complexity and emergent patterns can result from simple rules.
From one tick of the game to the next, a live cell with fewer than two or more than three live neighbors dies. A live cell with two or three neighbors lives on, and a dead cell with exactly three live neighbors becomes a live cell.

Game of Life

Each agent in this model begins with a random direction, but is constantly shifting its direction to be closer to the average of its closest neighbors. Eventually, all agents will be moving in roughly the same direction.

Flocking

This model extends the 2d flocking model into 3 dimensions, using Three.js instead of Flocc renderers to visualize the emergence of flocking behavior.
Like many other models, the environment is bounded such that agents that go outside of the cube will emerge on the other side.

Flocking (3d)

The Schelling Segregation model (1971) demonstrates how group-level patterns can emerge that are disconnected from the intentions of individuals. In this model, agents are represented by individual red or blue pixels on a <code>Terrain</code>. An agent belonging to each group wants to be surrounded by more of its own type than the other, and will move to an empty cell in the environment if it is in the minority.
These agents do not exhibit aversion to other types of agents per se, and only want to find a position in a non-minority group, but the resulting movements deterministically result in a segregated environment — an illustration of how systems-level effects (i.e. systemic bias) can arise unintuitively from lower-level behavior.

Segregation

If you&#8217;ve ever been on a freeway at rush hour, you know the stop-and-start nature of heavy traffic. In this model, an agent with an empty &#8216;road&#8217; in front of it will attempt to accelerate until it gets close to another agent, at which point it decelerates (i.e. slamming on the brakes). Watch how this behavior sends &#8216;waves&#8217; backwards to the agents behind it, causing &#8216;traffic jams.&#8217; Under the right conditions, agents are able to separate such that traffic becomes smooth, but other environments are plagued by disruptive, halting motion.

Traffic

The fire model simulates how fire spreads through a forest. It starts with a wave of trees simultaneously catching fire, and with every tick of the simulation, a tree with at least one neighboring tree (to the north, south, east, or west) on fire will also catch fire.
The PERCENT_FULL parameter determines how crowded the forest is. Set to 1.0 (100% full), every cell in the grid will be occupied by a tree, and the fire will deterministically sweep across the entire forest. However, around 0.59 (59% full), there is a threshold, and value below 0.59 will rarely result in the entire forest being burned, while above 0.59 the entire forest will almost always catch fire.
Read more at https://en.wikipedia.org/wiki/Forest-fire_model

Forest Fire

Cellular automata are discrete models that usually take place on a grid (or lattice), with cells that take action based on their own state (for example, &#8220;on&#8221; or &#8220;off&#8221;) and the state of nearby cells. The Game of Life is an example of a cellular automata. Elementary cellular automata are a subclass that takes place on a 1-dimensional grid (a single row).
Reading from top to bottom, each visualized row represents the state of the cells over time. Each rule from 0 to 255 represents the action a cell will take given its own state and that of its immediate neighbors. For example, in rule 0, every cell will turn off no matter what. In Rule 184, the &#8220;traffic rule,&#8221; the number of cells in each state always remains the same, but sends &#8216;waves&#8217; throughout the lattice. With random starting states, rules 30, 90, 110 exhibit unpredictable, chaotic behavior. Try other rules below!

Elementary Cellular Automata

This intentionally abstract simulation depicts the dynamics of a network tending toward the mean.
Each agent is connected to its closest geographic neighbors, and possibly a few other random agents in the network, depending on the value of REWIRING_PROBABILITY. With each tick of the simulation, an agent that is larger than its neighbor will transfer some of its size to the neighbor, shrinking the agent and growing the neighbor. With enough time, the network reaches stasis.
Try adjusting the parameters to see how it affects the network dynamics.

Network Diffusion

This implementation of Joshua Epstein&#8217;s <a href="https://www.pnas.org/content/99/suppl_3/7243, 2002">Civil Violence Model</a> uses two subclasses of the Agent base class — Civilian and Cop (see the link above for simulation parameters).
In this visualization, using the CanvasRenderer, Cops are represented by the letter c. Civilians are represented by a period, 0, or *. Every turn, some Civilians will go from inactive to active, depending on their own threshold and the exogenous legitimacy (i.e. of the government). However, a Cop who is next to an active Civilian will arrest them, visualized by 0 for a number of terms set by the 
jail term. By updating parameters for legitimacy, jail term, % full, and % Cops, different dynamics will arise, from a peaceful population to periodic outbursts to near-constant anarchy.

Civil Violence

A Voronoi diagram divides a plane into a set of non-overlapping regions, where each region is the set of points closest to a given point in that region.
One method of deriving a Voronoi diagram from a set of points on a plane is through the emergent movement of (many more) random vectors on the plane. In this example, 25 random points are placed on the plane to divide it into regions. Then, 1,000 random vectors (points with a random directional velocity) are placed and begin moving until they are the same distance from the two nearest points. Then, they stop. The resulting accumulation of stationary vectors describes the boundaries of the Voronoi diagram.

Emergent Voronoi

Based on the model described by Rudy Rucker in &#8220;Artificial Life Lab&#8221; and implemented in NetLogo, this model simulates a voting process and the emergence of polarized opinions. Each agent samples its neighbors within RADIUS and, if there&#8217;s a clear super-majority with a different vote from the agent, it will change its vote (i.e. bowing to social pressure).
By expanding the RADIUS parameter, regions of agents with the same opinion become larger and more defined.

Flocc

Examples