Friends of Tarok, Lord of Guac Chapter I: Party Fortress: Technical Post-Mortem

The highrise interface
The highrise interface

Last weekend Fei, Dan, and I put on our first Party Fortress party.

Fei and I have been working with social simulation for awhile now, starting with our Humans of Simulated New York project from a year ago.

Over the past month or two we've been working on a web system, "highrise" to simulate building social dynamics.

The goal of the tool is to be able to layout buildings and specify the behaviors of its occupants, so as to see how they interact with each other and the environment, and how each of those interactions influences the others.

Beyond its practical functionality (it still has a ways to go), highrise is part of an ongoing interest in simulation and cybernetics. Simulation is an existing practice that does not receive as much visibility as AI but can be just as problematic. It seems inevitable that it will become the next contested space of technological power.

highrise is partly a continuation of our work with using simulation for speculation, but whereas our last project looked at the scale of a city economy, here we're using the scale of a gathering of 10-20 people. The inspiration for the project was hearing Dan's ideas for parties, which are in many ways interesting social games. By arranging a party in a peculiar way, either spatially or socially, what new kinds of interactions or relationships emerge? Better yet, which interactions or relationships that we've forgotten start to return? What relationships that we've mythologized can (re)emerge in earnest?

highrise was the engine for us to start exploring this, which manifested in Party Fortress.

I'll talk a bit about how highrise was designed and implemented and then the living prototype Party Fortress.

Buildings: Floors and Stairs

First we needed a way to specify a building. We started by reducing a "building" to just floors and stairs, so we needed to develop a way to layout a building by specifying floor plans and linking them up with stairs.

Early sketches of highrise
Early sketches of highrise

We wanted floor plans to be easily specified without code, so developing some simple text structure seemed like a good approach. The first version of this was to simply use numbers:


Here 0 is empty space, 1 is walkable, and 2 is an obstacle. In the example above, each 2D array is a floor, so the complete 3D array represents the building. Beyond one floor it gets a tad confusing to see them stacked up like that, but this may be an unavoidable limitation of trying to represent a 3D structure in text.

Note that even though we can specify multiple floors, we don't have any way to specify how they connect. We haven't yet figured out a good way of representing staircases in this text format, so for now they are explicitly added and positioned in code.


A floor plan isn't enough to properly represent a building's interior - we also needed a system for specifying and placing arbitrary objects with arbitrary properties. To this end we put together an "object designer", depicted below in the upper-left hand corner.

The object designer
The object designer

The object designer is used to specify the footprint of an object, which can then be placed in the building. When an object is clicked on, you can specify any tags and/or key-value pairs as properties for the object (in the upper-right hand corner), which agents can later query to make decisions (e.g. find all objects tagged food or toilet).

Objects can be moved around and their properties can be edited while the simulation runs, so you can experiment with different layouts and object configurations on-the-fly.

Expanding the floor plan syntax

It gets annoying to need to create objects by hand in the UI when it's likely you'd want to specify them along with the floor plan. We expanded the floor plan syntax so that, in addition to specifying empty/walkable/obstacle spaces (in the new syntax, these are '-', ' ', and '#' respectively), you can also specify other arbitrary values, e.g. A, B, , etc, and these values can be associated with object properties.

Now a floor plan might look like this:

    "#.#.#. .#.#.#",
    "#. .#.   . .#",
    "#. .#.#.#. .#",
    "#.A. . . . .#",
    "#. .#.#.#.#.#",

And when initializing the simulation world, you can specify what A is with an object like this:

    'A': {
        'tags': ['food'],
        'props': {
            'tastiness': 10

This still isn't as ergonomic as it could be, so it's something we're looking to improve. We'd like it so that these object ids can be re-used throughout the floor plan, e.g:

    "#.#.#. .#.#.#",
    "#. .#.   .A.#",
    "#. .#.#.#. .#",
    "#.A. . .A. .#",
    "#. .#.#.#.#.#",

so that if there are identical objects they don't need to be repeated. Where this becomes tricky is if we have objects of footprints larger than one space, e.g.:

    "#.#.#. .#.#.#",
    "#. .#.   . .#",
    "#. .#.#.#. .#",
    "#.A.A. . . .#",

Do we have three adjacent but distinct A objects or one contiguous one?

This ergonomics problem, in addition to the stair problem mentioned earlier, means there's still a bit of work needed on this part.

(Spatial) Agents

The building functionality is pretty straightforward. Where things start to teeter (and get more interesting) is with designing the agent framework, which is used to specify the inhabitants of the building and how they behave.

It's hard to anticipate what behaviors one might want to model, so the design of the framework has flexibility as its first priority.

There was the additional challenge of these agents being spatial; my experience with agent-based models has always hand-waved physical constraints out of the problem. Agents would decide on an action and it would be taken for granted that they'd execute it immediately. But when dealing with what is essentially an architectural simulation, we needed to consider that an agent may decide to do something and need to travel to a target before they can act on their decision.

So we needed to design the base Agent class so that when a user implements it, they can easily implement whatever behavior they want to simulate.

Decision making

The first key component is how agents make decisions.

The base agent code is here, but here's a quick overview of this structure:

  • There are two state update methods:
    • entropy, which represents the constant state changes that occur every frame, regardless of what action an agent takes. For example, every frame agents get a bit more hungry, a bit more tired, a bit more thirst, etc.
    • successor, which returns the new state resulting from taking a specific action. This is applied only when the agent reaches its target. For example, if my action is eat, I can't actually eat and decrease my hunger state until I get to the food.
  • actions, which returns possible actions given a state. E.g. if there's no food at the party, then I can't eat
  • utility, which computes the utility for a new state given an old state. For example, if I'm really hungry now and I eat, the resulting state has lower hunger, which is a good thing, so some positive utility results.
    • Agents use this utility function to decide what action to take. They can either deterministically choose the action which maximizes their utility, or sample a distribution of actions with probabilities derived from their utilities (i.e. such that the highest-utility action is most likely, but not a sure bet).
    • This method also takes an optional expected parameter to distinguish the use of this method for deciding on the action and for actually computing the action's resulting utility. In the former (deciding), the agent's expected utility from an action may not actually reflect the true result of the action. If I'm hungry, I may decide to eat a sandwich thinking it will satisfy my hunger. But upon eating it, I might find that it actually wasn't that filling.
  • execute, which executes an action, returning the modified state and other side effects, e.g. doing something to the environment.


Agents also can have an associated Avatar which is their representation in the 3D world. You can hook into this to move the agent and know when it's reached it's destination.

Agent movement
Agent movement

Multi-floor movement was handled by standard A* pathfinding:

A* pathfinding (Wikipedia)
A* pathfinding (Wikipedia)

Each floor is represented as a grid, and the layout of the building is represented as a network where each node is a floor and edges are staircases. When an agent wants to move to a position that's on another floor, they first generate a route through this building network to figure out which floors they need to go through, trying to minimize overall distance. Then, for each floor, they find the path to the closest stairs and go to the next floor until they reach their target.

Building network example
Building network example

There are some improvements that I'd really like to make to the pathfinding system. Currently each position is weighted the same, but it'd be great if we held different position weights for each individual agent. With this we'd be able to represent, for instance, subjective social costs of spaces. For example, I need to go to the bathroom. Normally I'd take the quickest path there but now there's someone I don't want to talk there. Thus the movement cost of those positions around that person are higher to me than they are to others (assuming everyone else doesn't mind them), so I'd take a path which is longer in terms of physical distance, but less imposing in terms of overall cost when considering this social aspect.

Subjective path weight example
Subjective path weight example

Those are the important bits of the agent part. When we used highrise for Party Fortress (more on that below), this was enough to support all the behaviors we needed.

Party Fortress

Since the original inspiration for highrise was parties we wanted to throw a party to prototype the tool. This culminated in a small gathering, "Party Fortress" (named after Dwarf Fortress), where we ran a simulated party in parallel to the actual party, projected onto a wall.


We wanted to start by simulating a "minimum viable party" (MVP), so the set of actions in Party Fortress are limited, but essential for partying. This includes: going to the bathroom, talking, drinking alcohol, drinking water, and eating.

The key to generating plausible agent behavior is the design of utility functions. Generally you want your utility functions to capture the details of whatever phenomena you're describing (this polynomial designer tool was developed to help us with this).

For example, consider hunger: when hunger is 0, utility should be pretty high. As you get hungry, utility starts to decrease. If you get too hungry, you die. So, assuming that our agents don't want to die (every simulation involves assumptions), we'd want our hunger utility function to asymptote to negative infinity as hunger increases. Since agents use this utility to decide what to do, if they are really, really hungry they will prioritize eating above all else since that will have the largest positive impact on their utility.

Hunger utility function
Hunger utility function

So we spent a lot of time calibrating these functions. The more actions and state variables you add, the more complex this potentially gets, and makes calibration much harder. We're still trying to figure out a way to make this a more streamlined process involving less trial-and-error, but one helpful feature was visualizing agents' states over time:

Agent state charts
Agent state charts


One challenge with spatial agents is that as they are moving to their destination, they may suddenly decide to do something else. Then, on the way to that new target, they again may decide to something else. So agents can get stuck in this fickleness and never actually accomplish anything.

To work around this we incorporated a commitment variable for each agent. It feels a bit hacky, but basically when an agent decides to do something, they have some resolve to stick with it unless some other action becomes overwhelmingly more important. Technically this works out to mean that whatever action an agent does has its utility artificially inflated (so it's more appealing to continue doing it) until they finally execute it or the commitment wears off. This could also be called stubbornness.


Since conversation is such an important part of parties we wanted to model it in higher fidelity than the other actions. This took the form of having varying topics of conversation and bestowing agents with preferences for particular topics.

We defined a 2D "topic space" or "topic matrix", where one axis is how "technical" the topic is and the other is how "personal" the topic is. For instance, a low technical, low personal topic might be the weather. A high technical but low personal topic might be the blockchain.

Conversation topic space
Conversation topic space

Agents don't know what topic to talk about with an agent they don't know, but they a really basic conversation model which allows them to learn (kind of, this needs work). They'll try different things and try to gauge how the other person responds, and try to remember this.

Social Network

As so far specified, our implementation of agents don't capture, explicitly at least, the relationships between individual agents. In the context of a social simulation this is obviously pretty important.

For Party Fortress we implemented a really simple social network so we could represent pre-existing friendships and capture forming ones as well. The social network is modified through conversation and the valence and strength of modification is based on what topics people like. For example, if we talk about a topic we both like, our affinity increases in the social graph.


It's not very interesting to watch the simulation with no other indicators of what's happening. These are people we're supposed to be simulating and so we have some curiosity and expectations about their internal states.

We implemented a narrative system where agents will report what exactly their doing in colorful ways.

Agents talking and thinking
Agents talking and thinking

Closing the loop

Our plan for the party was to project the simulation on the wall as the party went on. But that introduces an anomaly where our viewing of the simulation may influence our behavior. We needed the simulation itself to capture this possibility - so we integrated a webcam stream into the simulation and introduced a new action for agents: "gawk". Now they are free to come and watch us, the "real" world, just as we can watch them.

Agents talking and thinking
Agents talking and thinking

We have a few other ideas for "closing the loop" that we weren't able to implement in time for Party Fortress I, such as more direct communication with simulants (e.g. via SMS).


We hosted Party Fortress at Prime Produce, a space that Dan has been working on for some time.

We had guests fill out a questionnaire as they arrived, designed to extract some important personality features. When they submitted questionnaire a version of themselves would appear in the simulation and carry on partying.


There were surprisingly several moments of synchronization between the "real" party and the simulated one. For instance, people talking or eating when the simulation "predicted" it. Some of the topics that were part of the simulation came up independently in conversation (most notably "blockchain", but that was sort of a given with the crowd at the party). And of course seeing certain topics come up in the simulation spurned those topics coming up outside of it too.

The Party
The Party

Afterwards our attendees had a lot of good feedback on the experience. Maybe the most important bit of feedback was that the two parties felt too independent; we need to incorporate more ways for party-goers to feel like they can influence the simulated party and vice versa.

It was a good first step - we're looking to host more of these parties in the future and expand highrise so that it can encompass weirder and more bizarre parties.

Social Simulation Components: Social Contagion

The idea of syd is that it will be geared towards social simulation - that is, modeling systems that are driven by humans (or other social beings) interacting. To support social simulations syd will include some off-the-shelf models that can be composed to define agents with human-ish behaviors.

One category of such models are around the propagation and mutation of ideas and behaviors ("social contagion"). Given a population of agents with varying cultural values, how do these values spread or change over time? How do groups of individuals coalesce into coherent cultures?

What follows are some notes on a small selection of these models.

Sorting & peer effects

Two primary mechanisms are sorting, or "homophily", the tendency to associate with similar people, and peer effects, the tendency to become more like people we are around often.

Schelling's model of segregation may be the most well-known example of a sorting model, where residents move if too many of their neighbors aren't like them.

Schelling's model source
Schelling's model source

A very simple peer effect model is Granovetter's model. Say there is a population of $N$ individuals and $n$ of them are involved in a riot. Each individual in the population has some threshold $T_i$; if over $T_i$ people are in the riot, they'll join the riot too. Basically this is a bandwagon model in which people will do whatever others are doing when it becomes popular enough.

Granovetter's model does not incorporate the innate appeal of the behavior (or product, idea, etc), just how many people are participating in it. One could imagine though that the riot looks really exciting and people join not because of how many others are already there, but because they are drawn to something else about it.

A simple extension of Granovetter's model, the Standing Ovation model, captures this. The behavior in question has some appeal or quality $Q$. We add some additional noise to $Q$ to get the observed "signal" $S = Q + \epsilon$. This error allows us to capture a bit of the variance in perceived appeal (some people may find it appealing, some people may not, the politics around the riot may be very complex, etc). If $S > T_i$, then person $i$ participates. After assessing the appeal, those who are still not participating then have the Granovetter's model applied (they join if enough others are participating).

There are further extensions to the Standing Ovation model. You could say that an agent's relationships affects their likelihood to join, e.g. if I see 100 strangers rioting I may not care to join, but if 10 of my friends are then I will.

Axelrod's culture model

Axelrod's culture model is a simple model of how a "culture" (a population sharing many traits, beliefs, behaviors, etc) might develop.

  • each individual is described by a vector of traits
  • the population is placed in some space (e.g. a grid or a social network)
  • an individual interacts with a neighbor with some probability based on their trait similarity
  • if they interact, pick one trait and match with the neighbor's
A social network version of Axelrod's cultural disseminiation model source
A social network version of Axelrod's cultural disseminiation model source

This model can be extended with a consistency rule: sometimes the value of one trait is copied over to another trait (in the same individual), this models the two traits becoming consistent. For example, perhaps the changing of one belief causes dependent beliefs to change as well, or requires beliefs it's dependent on to change.

Traits can also randomly change as well due to "error" (imperfect imitation or communication) or "innovation".

Modeling opinion flow with Boids

A really interesting idea is to apply the Boids flocking algorithm to the modeling of idea dissemination (if you're unfamiliar with Boids, Craig Reynolds explains here).

The original boids source
The original boids source

Here agents form a directed graph (e.g. a social network), where edges have two values: frequency of communication and respect one agent holds for the other. For each possible belief, agents have an alignment score which can have an arbitrary scale, e.g. -3 for strong disbelief to 3 for strong belief.

The agent feels a "force" that causes them to change their own alignment. This force is the sum of force from alignment, force from cohesion, and force from separation.

  • force from alignment is computed by the average alignment across all agents - this is the perceived majority opinion.
  • force from cohesion: each agent computes the average alignment felt by their neighbors they respect (i.e. respect is positive), weighted by their respect for those neighbors.
  • force from separation: like the force from cohesion, but computed across their neighbors they disrespect (i.e. respect is negative), weighted by their respect for those neighbors.

This force is normalized and is used to compute a probability that the individual changes their alignment one way or the other. We specify some proportionality constant $\alpha$ which determines how affected an agent is by the force. For force $F$ the probability of change of alignment is just $\alpha F$. It's up to the implementer how much an agent's alignment changes.


A meme is an idea/belief/etc that behaves according to the rules of memetics, which models the spread of ideas by applying evolutionary concepts:

  1. phenotypes & alleles
    • the "offspring" of a meme vary in their "appearance"
    • a meme contains characteristics ("alleles"), some of which are transmitted to their child
    • variability in allele combinations is responsible for variability at the phenotypic level
  2. mutation
    • idea mutation may be random
    • or it may happen for a reason; ideas can change — to solve problems, for instance (these mutations are essentially innovations advocated by a change agent)
  3. selection
    • some ideas are more likely to survive than others
    • an idea's survival is based on how "fit" it is
    • an idea's measure of fitness is the likelihood of its offspring surviving long enough to produce their own offspring, compared to other memes
  4. Lamarckian properties
    • unlike biological evolution, members can be modified, activated or deactivated within a generation (people can adapt their ideas to deal with new information, for example)
  5. drift
    • if multiple finite-sized populations exist, beginning w/ the same set of initial conditions & operate according to the same mechanisms/constraints, completely different sets of ideas can emerge b/w the populations
    • this drift is due to sampling error when a parent meme produces offspring (random allele heritage)

Memetic transmission may be horizontal (intra-generational) and/or vertical (intergenerational).

The primary mechanism for memetic transmission is imitation, but other mechanisms include social learning and instruction.

Note that the transmission of an idea is heavily dependent on its own characteristics.

For example, there are some ideas that have "mutation-resistant" qualities, e.g. "don't question the Bible" (though that does not preclude them from mutation entirely).

Some ideas also have the attribute of "proselytism"; that is part of the idea is the spreading of the idea.

The Cavalli-Sforza Feldman model is a memetics model describing how a population of beliefs evolve over time.

There is a transmission function:

$$ p = 1 - |1 - g|^{n \mu_t} $$


  • $p$ is the probability that an individual's belief state will be transformed (i.e. imitate another's) after $n$ contacts
  • $g$ is the probability of transformation after each contact
  • $\mu_t$ is the proportion of people the individual can come into contact with who already have the target belief state

There's also a selection function:

$$ \mu_t' = \frac{\mu_t (1+s)}{1 + s \mu_t} $$


  • $\mu_t'$ is the proportion of beliefs that survive selection for a single generation
  • $\mu_t$ is the proportion of beliefs before selection
  • $s$ is the degree of fitness

The models presented so far make no distinction between public and private beliefs, but it's very clear that people often publicly conform to ideas that they may not really feel strongly about. This phenomenon is called pluralistic ignorance: "situations where a majority of group member privately reject a norm but assume (incorrectly) that most others accept it".

It's one thing to publicly conform to ideas you don't agree with, but people often go a step further and try to enforce others to comply to these ideas. Why is that?

The "illusion of transparency" refers to the tendency where people believe that others can read more about their internal state than they actually can. Maybe something embarrassing happened and you feel as if everyone knows, even if no one possibly could. In the context of pluralistic ignorance, this tendency causes people to feel as though others can see through their insincere alignment with the norm, so they take additional steps to "prove" their conviction.

The Emperor’s Dilemma: A Computational Model of Self‐Enforcing Norms proposes the following model for this phenomena:

  • each agent $i$ has a binary private belief $B_i$ which can be 1 (true believer) or -1 (disbeliever).
  • true enforcement is when a true believer/disbeliever enforces others to comply/oppose
  • false enforcement is when a false believer enforces others to comply

An agent $i$ choice to comply with the norm is $C_i$. If $C_i=1$ the agent chooses to complex, otherwise $C_i=-1$. This choice depends on the strength of the agent's convictions $0 < S \leq 1$.

A neighbor $j$'s enforcement of the norm is represented as $E_j=1$; if they enforce deviance instead then $E_j=-1$. Thus we can compute $C_i$ as:

$$ C_i = \begin{cases} -B_i & \text{if } \frac{-B_i}{N_i} \sum_{j=1}^{N_i} E_j > S_i \\ B_i & \text{otherwise} \end{cases} $$

We assume that true believers ($B_i = 1$) have maximal conviction ($S_i = 1$) and so are resistant to neighbors enforcing deviance.

The enforcement of an agent $i$ is computed:

$$ E_i = \begin{cases} -B_i & \text{if } (\frac{-B_i}{N_i} \sum_{j=1}^{N_i} E_j > S_i + K) \land (B_i \neq C_i) \\ +B_i & \text{if } (S_i W_i > K) \land (B_i = C_i) \\ 0 & \text{otherwise} \end{cases} $$

where $0 < K < 1$ is an additional cost of enforcement for those who also comply (it is $K$ more difficult to get someone who does not privately/truly align with the belief to enforce it).

$W_i$ is the need for enforcement, which is the proportion of agent $i$'s neighbors whose behavior does not confirm with $i$'s beliefs $B_i$:

$$ W_i = \frac{1- (B_i/N_i) \sum_{j=1}^{N_i} C_j}{2} $$

Agents can only enforce compliance or deviance if they have complied or deviated, respectively.

The model can be extended by making it so that true disbelievers can be "converted" to true believers (i.e. their private belief changes to conform to the public norm).


Social Simulation Components: Cultural Values

The agents in syd will need to be parameterized in some way that meaningfully affects their behavior. Another way to put this is that the agents need some values that guide their actions. In Humans of Simulated New York Fei and I defined individuals along the axes of greed vs altruism, lavishness vs frugality, long-sightedness vs short-sightedness, and introversion vs extroversion. The exact configuration of these values are what made an agent an individual: a lavish agent would spend more of their money, an extroverted agent would have a larger network of friends (which consequently made finding a job easier), greedy agents would pay their employees less, and so on.

HOSNY value dimensions
HOSNY value dimensions

The dimensions we used aren't totally comprehensive. There are many aspects of human behavior that they don't encapsulate. Fortunately there is quite a bit of past work to build on - there have been many past attempts to inventory a value spectrums that defines and distinguishs cultures. The paper A Proposal for Clustering the Dimensions of National Culture (Maleki, A., de Jong, M, 2014) neatly catalogues these previous efforts and proposes their own measurements as well.

The authors propose the following cultural dimensions:

  • individualism vs collectivism
  • power distance: "the extent to which hierarchical relations and position-related roles are accepted"
  • uncertainty avoidance: "to what extent people feel uncomfortable with certain, unknown, or unstructured situations"
  • mastery vs harmony: "competitiveness, achievement, and self-assertion versus consensus, equity, and harmony"
  • traditionalism vs secularism: "religiosity, self-stability, feelings of pride and, consistency between emotion felt and their expression vs secular orientation and flexibility"
  • indulgence vs restraint: "the extent to which gratification of desires and feelings is free or restrained"
  • assertiveness vs tenderness: "being assertive and aggressive versus kind and tender in social relationships"
  • gender egalitarianism
  • collaborativeness: "the spirit of 'team-work'"

We can (imprecisely) map the dimensions we used in HOSNY to these:

  • greed vs altruism -> individualism vs collectivism and collaborativeness
  • lavishness vs frugality -> indulgence vs restraint
  • long-sightedness vs short-sightedness -> indulgence vs restraint
  • introversion vs extroversion -> assertiveness vs tenderness (?)

It doesn't feel very exact though.

All of the previously defined dimensions are worth a look too:

Simulation Designer Sketches

The other day I wrote a bit about the backend architecture of syd; here I'll talk a bit about the tentative design plans. While the backend part of syd is supposed to make writing agent-based simulations easier, you still need to know implement these simulations with code. A typical process for developing such simulations involves design phases, e.g. pen-and-paper sketches or diagramming with flowcharting software where the high-level structure of the system is laid out. Causal loop diagrams are often used in this way.

A causal loop diagram
A causal loop diagram

Ideally the process of designing and sketching this high-level system architecture is the same as implementing the system simulation. This isn't a new idea; it's the approach software like Vensim uses. The point of syd, however, is to appeal to people with little to no systems thinking background; existing systems modeling software is intended for professionals in academia and industry in addition to being closed source and expensive.

Node-based interface

The general idea is to use a node-based interface. With node-based interfaces you compose causal graphs which are a natural and intuitive representation of systems. As an added bonus, if you design it so that nodes can be composed of other nodes, the interface lends itself to modularity as well.

A node-based interface (quadrigram)
A node-based interface (quadrigram)

System and agent views

In the syd interface there are two view levels:

  • the system level (i.e. the "environment" or the "world"). This works like conventional system dynamics software such as Vensim; i.e. you can define stocks and flows and connect them together.
  • the agent level. In this view you design the internals of a type of agent (e.g. a person, a firm, etc).

Since the system level view is quite similar to conventional system dynamics software (and also because I haven't fully thought it through yet) I won't go into much detail there. Basically the system level supports the creation of stocks (quantities), flows (changes in quantities that can link between stocks), and outputs (e.g. graphs and other visualizations). This gif of TRUE gives a good sense of it:


For example, a flow may be some arbitrary function that other outputs and inputs can link to.

Arbitrary function node
Arbitrary function node

You can also take a bunch of flows and stocks and group them into a module node, which "black boxes" the internals so you avoid spaghetti:

Node-based spaghetti
Node-based spaghetti

You can also visualize aggregate statistics for agent types as well.

Agent aggregations
Agent aggregations

The system view is also where you spawn populations of agents.

Designing agents

Agents are defined as types (e.g. a person, or a firm, or a car driver, etc) and are composed of state variables and behaviors.

State variables have a name and may be of a particular type (e.g. discrete/categorical or continuous, perhaps collections or references as well). Depending on its type it may have additional parameters that can be set (e.g. if it is a continuous state variable, it may have minimum and maximum values).

State variables may be instantiated with a hardcoded value, which is identical across all agents of that type, or they may be instantiated with some function (e.g. a random distribution or a distribution learned from data), which may cause its value to vary for each individual agent. Note that a limitation here is that at instantiation state variables are treated as independent; i.e. we can't (yet) instantiate them using conditional distributions (conditioned on other state variables, for instance).

State variable instantiations
State variable instantiations

State variables are changed over the course of the simulation through behaviors. Behaviors are isolated components that are attached to an agent.

A behavior
A behavior

So far behaviors are black-boxes that read from some state variables and write to other state variables. A warning is raised if a required state variable is not present; perhaps there should be an option to map the expected state variable name to an existing state variable (e.g. if a behavior expects the state variable foo but my agent has bar, I can tell the behavior to use bar instead).

A warning will also be raised if there are multiple behaviors that need to write to the same state variable. So far I haven't thought of a way for the user to easily specify how such write conflicts should be resolved.

Behavior warning
Behavior warning

A special kind of behavior is a metabehavior, which, instead of modulating state variable values, adds or removes other behaviors based on conditions. I haven't yet figured out the best way to represent this in the interface.

Final notes

There are a few additional features that would be great to include:

  • export models as JSON (or some other nice interchange format)
  • hierarchical simulations; i.e. take another model and include it as a node in another simulation. For instance, maybe you have a simulation about the internal workings of a company (modeling employees and so on) and you want to use that as a sub-simulation in a model of a national economy.
  • level-of-depth (LOD) and multiscale simulations

These extra features and some other aspects of the interface as described here (especially agent behaviors) require re-architecting of some of the backend, so I don't know when we'll be able to prototype the interface. I'm not totally confident that this approach will be general/flexible enough for more complex simulations, but we'll see when we start to prototype and use it.

Big Simulation Architecture

Game of Life with syd
Game of Life with syd

I've been interested in extending the work on Humans of Simulated New York into a more general agent-based simulation framework, one that is both accessible to non-technical audiences and powerful enough for more "professional" applications as well. We are all so ill-equipped to contend with the obscenely complex systems we're a part of, and we end up relying on inadequate (and often damaging) heuristics that cause us to point blame at parties that either have little do with our problems or are similarly victimized by them. Maybe if we had tools which don't require significant technical training to help us explore these causal rat's nests, to make this insane interconnectedness more intuitive and presentable, we could think through and talk about these wicked problems more productively.

Recently Fei and I started working on a project, tentatively called "system designer" (syd for short), which we hope will provide a foundation for that kind of tooling. syd is currently functional but its features are provisional - I've included some examples of visualizations built on top of it, demoing simple models, although it is capable of more sophisticated agent-based models as well.

From an engineering perspective, the goal is to make it easy to write agent-based simulations which may have massive amounts of computationally demanding agents without having to deal with the messiness of parallel and distributed computing.

From a design perspective, the goal is to provide interfaces that make defining, running, visualizing, and analyzing such simulations an exploratory experience, almost like a simulation video game (e.g. SimCity).

In both the design and engineering cases there are many interesting challenges.

I'm going to discuss the engineering aspects of the project here and how syd is approaching some of these problems (but bear in mind syd is still in very early stages and may change significantly). At another point I'll follow-up with more about the design and interface aspects (this is something we're still sketching out).

syd is built on top of aiomas which handles lower-level details like inter-node communication, so I won't discuss those here.

(note that at time of writing, not all of what's discussed here has been implemented yet)

3D Schelling model with syd
3D Schelling model with syd

The demands of simulation

If you're conducting a relatively lightweight simulation, like cellular automata, in which agents are composed of a few simple rules, you can easily run it on a single machine, no problem. The update of a single agent takes almost no time.

Unfortunately, this approach starts to falter as you get into richer and more detailed simulations. In our first attempts at Humans of Simulated New York, Fei and I designed the agents to be fairly sophisticated - they would have their own preferences and plan out a set of actions for each day, re-planning throughout the day as necessary. Even with our relatively small action space (they could work, look for work, relax, or visit a doctor), this planning process can take quite awhile, especially when it's executed by hundreds or thousands of agents.

Here you could turn to parallel or distributed methods: you'd take your population of agents, send them to a bunch of different threads or processes or computers (generally referred to as "multi-node" architecture), and then update them in parallel. For example, if you run your simulation across two machines instead of just one, you can almost double the speed of your simulation.

Normally to convert a single-node simulation to a multi-node one you'd have to change up your code to support communication across nodes and a laundry list of other things, but syd abstracts away the difference. You simply instantiate a ComputeSubstrate and you pass in either a single host or a list of hosts. If you pass a single host, the simulation runs in the local process; if a list of hosts, syd transparently runs it as a multi-node simulation:

from syd import ComputeSubstrate

single_node = ComputeSubstrate(('localhost', 8888))

multi_node = ComputeSubstrate([
        ('localhost', 8888),
        ('localhost', 8889),
        ('', 8888),
        ('', 8889)])

Determining where agents go

That's great, but it doesn't come for free. Consider a simulation in which each agent must consult a few other agents before deciding what to do. For a single-node (here I'm using a "node" to refer to a single serial process) simulation this communication would happen quickly - all agents are in the same process so there's basically no overhead to get data from one another.

As soon as we move to the multi-node case we have to worry about the overhead that network communication introduces. The computers we distribute our population across could be on separate continents, or maybe we just have a terrible internet connection, and there may be considerable lag if an agent on one node needs a piece of data from an agent on another node. This network overhead can totally erase all speed gains we'd get from distributing the simulation.

The typical way of managing this network overhead is to be strategic about how agents are distributed across the nodes. For example, if we're simulating some kind of social network, maybe agents really only communicate with their friends and no one else. In this case, we'd want to put groups of friends in the same process so they don't have to go over the network to communicate, and we'd effectively eliminate most of the network communication.

The problem here (maybe you're starting to see a pattern) is that there is no one distribution strategy that works well for all conceivable agent-based simulations. It really depends on the particular communication patterns that happen within the simulation. In the literature around engineering these agent-based systems you'll see mention of "spheres of influence" and "event horizons" which determine how to divide up a population across nodes. Agents that are outside of each other's spheres of influence or beyond each other's event horizons are fair game to send to different nodes. Unfortunately what exactly constitutes a "sphere of influence" or "event horizon" varies according to the specifics of your simulation.

In syd, if you create a multi-node substrate you can also specify a distribution strategy (a Distributor object) which determines which agents are sent to which nodes. So far there are only two strategies:

  • syd.distributors.RoundRobin: a naive round robin strategy that just distributes agents sequentially across nodes.
  • syd.distributors.Grid2DPartition: if agents live in a grid, this recursively bisects the grid into "neighborhoods" so that each neighborhood gets its own node. This is appropriate when agents only communicate with adjacent agents (e.g. cellular automata). Network communication still happens at the borders of neighborhoods, but overall network communication is minimized.

You can also define your own Distributor for your particular case.

from syd import ComputeSubstrate, distributors

multi_node = ComputeSubstrate([
        ('localhost', 8888),
        ('localhost', 8889),
        ('', 8888),
        ('', 8889)],

There is yet another problem to consider - in some simulations, agents may be mobile; for example, if we're simulating a grid world, we may distribute agents according to where in the grid they are (e.g. with the Grid2DPartition distributor), but what if they move to another part of the grid? We might want to move them to the node that's running that part of the grid. But if this happens a lot, now we've introduced a ton of overhead shuffling agents from node to node.

As another example - if the topology of the simulation is a social network instead of a grid, such that they are communicating most with their friends, what happens if those relationships change over time? If they become friends with agents on another node, should we re-locate them to that node? Again, this will introduce extra overhead.

I haven't yet come up with a good way of handling this.

Social network SIR model with syd
Social network SIR model with syd

Race conditions and simultaneous updates

There is yet another problem to consider. With a single-node simulation, agents all update their states serially. There is no fear of race conditions: we don't have to worry about multiple agents trying to simultaneously access a shared resource (e.g. another agent) and making conflicting updates or out-of-sync reads.

This is a huge concern in the multi-node case, and the way syd works around it is to separate agent updates into two distinct phases:

  • the decide phase, a.k.a. the "observation" or "read" phase, where the agent collects ("observes") the information it needs from other agents or the world and then decides on what updates to make (but it does not actually apply the updates).
  • the update phase, a.k.a. the "write" phase, where the agent applies all queued updates.

So in the decide phase, agents queue updates for themselves or for other agents as functions that mutate their state. This has the effect of agents simultaneously updating their states, as opposed to updating them in sequence.

Unfortunately, this structure does not entirely solve our problems - it is possible that there are conflicting or redundant updates queued for an agent, and those usually must be dealt with in particular ways. I'm still figuring out a good way to manage this.

Node failures

Another concern is the possibility of node failure. What if a machine dies in the middle of your simulation? All the agents that were on that machine will be lost, and the simulation will become invalid.

The approach that I'm going to try is to snapshot all agent states every $n$ steps (to some key-value store, ideally with some kind of redundancy in the event that one of those nodes fail!). If a node failure is detected, the simulation supervisor will automatically restart the simulation from the last snapshot (this could involve the spinning up of a new replacement machine or just continuing with one less machine).

Abstract the pain away

Long story short: parallel and distributed computing is hard. Fortunately for the end-user, most of this nasty stuff is hidden away. The plan is to include many "sensible defaults" with syd so that the majority of use-cases do not require, for example, the implementation of a custom Distributor, or any other messing about with the internals. You'll simply indicate that you want your simulation to run on multiple computers, and it'll do just that.


  • Antelmi, A., Cordasco, G., Spagnuolo, C. & Vicidomini, L. (2015). On Evaluating Graph Partitioning Algorithms for Distributed Agent Based Models on Networks. European Conference on Parallel Processing.
  • Bharti, R. (2016). HIVE - An Agent Based Modeling Framework. Master's Projects.
  • Chenney, S. (2001). Simulation Level-of-Detail. Game Developers Conference.
  • Chenney, S. Simulation Culling and Level-of-Detail. IEEE Computer Graphics and Applications.
  • Chenney, S., Arikan, O. & Forsyth, D. (2001). Proxy Simulations for Efficient Dynamics. EUROGRAPHICS.
  • Chris Rouly, O. (2014). Midwife: CPU cluster load distribution of Virtual Agent AIs. Complex, Intelligent and Software Intensive Systems (CISIS).
  • He, M., Ruan, H. & Yu, C. (2003). A Predator-Prey Model Based on the Fully Parallel Cellular Automata. International Journal of Modern Physics C.
  • Holcombe, M., Coakley, S., Kiran, M., Chin, S., Greenough, C., Worth, D., Cincotti, S., Raberto, M., Teglio, A., Deissenberg, C., van der Hoog, S., Dawid, H., Gemkow, S., Harting, P. & Neugart, M. (2013). Large-Scale Modeling of Economic Systems. Complex Systems, 22.
  • K. Bansal, A. (2006). Incorporating Fault Tolerance in Distributed Agent Based Systems by Simulating Bio-computing Model of Stress Pathways. Proceedings of SPIE, 6201.
  • K. Som, T. & G. Sargen, R. (2000). Model Structure and Load Balancing in Optimistic Parallel Discrete Event Simulation. Proceedings of the Fourteenth Workshop on Parallel and Distributed Simulation.
  • Kim, I., Tsou, M. & Feng, C. (2015). Design and implementation strategy of a parallel agent-based Schelling model. Computers, Environment and Urban Systems, 49(2015), pp. 30-41.
  • Kubalík, J., Tichý, P., Šindelář, R. & J. Staron, R. (2010). Clustering Methods for Agent Distribution Optimization. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews).
  • Lees, M., Logan, B., Oguara, T. & Theodoropoulos, G. (2004). Simulating Agent-Based Systems with HLA: The Case of SIM_AGENT - Part II. International Conference on Computational Science.
  • Logan, B. & Theodoropoulos, G. (2001). The Distributed Simulation of Multi-Agent Systems. Proceedings of the IEEE, 89(2).
  • Lysenko, M. & M. D'Souza, R. (2008). A Framework for Megascale Agent Based Model Simulations on Graphics Processing Units. Journal of Artificial Societies and Social Simluation, 11(4/10).
  • Márquez, C., César, E. & Sorribes, J. (2015). Graph-Based Automatic Dynamic Load Balancing for HPC Agent-Based Simulations. European Conference on Parallel Processing.
  • Navarro, L., Flacher, F. & Corruble, V. (2011). Dynamic Level of Detail for Large Scale Agent-Based Urban Simulations. Proceedings of 10th International Conference on Autonomous Agents and Multiagent Systems, pp. 701-708.
  • Oey, M., van Splunter, S., Ogston, E., Warnier, M. & M.T. Brazier, F. (2010). A Framework for Developing Agent-Based Distributed Applications. BNAIC 2010: 22rd Benelux Conference on Artificial Intelligence.
  • Pleisch, S. & Schiper, A. (2000). Modeling Fault-Tolerant Mobile Agent Execution as a Sequence of Agreement Problems. Reliable Distributed Systems.
  • Richiardi, M. & Fagiolo, G. (2014). Empirical Validation of Agent-Based Models.
  • Rousset, A., Herrmann, B., Lang, C. & Philippe, L. (2015). A communication schema for parallel and distributed Multi-Agent Systems based on MPI. European Conference on Parallel Processing.
  • Rubio-Campillo, X. (2014). Pandora: A Versatile Agent-Based Modelling Platform for Social Simulation. SIMUL 2014: The Sixth International Conference on Advances in System Simulation.
  • Scheutz, M. & Schermerhorn, P. (2006). Adaptive Algorithms for the Dynamic Distribution and Parallel Execution of Agent-Based Models. Journal of Parallel and Distributed Computing.
  • Sunshine-Hill, B. (2013). Managing Simulation Level-of-Detail with the LOD Trader. Motion in Games.
  • Tsugawa, S., Ohsaki, H. & Imase, M. (2012). Lightweight Distributed Method for Connectvity-Based Clustering Based on Schelling's Model. 26th International Conference on Advanced Information Networking and Applications Workshops.
  • Vecchiola, C., Grosso, A., Passadore, A. & Boccalatte, A. (2009). AgentService: A Framework for Distributed Multi-Agent System DEvelopment. International Journal of Computers and Applications, 31(3), pp. 204-210.
  • Dubitzky, W., Kurowski, K., & Schott, B, eds. (2012). Large-Scale Computing Techniques for Complex System Simulations.