Thought Experiments in the Browser

Brian Coffey
- San Francisco, CA

As data scientists, we work in concert with other members of an organization with the goal of making better decisions. This often involves finding trends and anomalies in historical data to guide future action. But in some cases, the best aid to decision-making is less about finding “the answer” in the data and more about developing a deeper understanding of the underlying problem. In this post we will focus another tool that is often overlooked: interactive simulations through the means of agent based modeling.

Simulating Agents

Agent based modeling involves the description of individual agents that interact with each other within an environment and seeing how their behaviours combine to produce macro-level system behaviours.

Agents can be modeled at whatever level seems natural to our understanding of the system: individual humans, client cohorts, departments, competing firms, computer programs or similar entities can all be agents. The environment may contain constraints or resources that influence and/or are influenced by the agents’ behaviors. The agents may interact directly or indirectly via the environment.

This conceptual framework allows you to model knowledge or assumptions at the smallest scale, where they can often be reasoned about more intuitively, rather than at the aggregate level. It also allows you to construct models that may not faithfully represent the system as it currently exists (e.g., more or fewer agents, different boundary conditions), but let you reason about the system and explore its complexities and possible outcomes.1

Agent based modeling is one of many possible frameworks for interactive simulation, but its particular benefits map well to business analysis and development at young, growing companies:

Its conceptual underpinnings are quickly grasped.It focuses model-making and model-describing attention at a micro level. It allows for observation and interaction with future projections at a macro level, with a greater focus on representing the underlying dynamics properly - i.e., getting the mental model right - than on pinpoint accuracy. Decision-making must happen quickly.The past is not always a good indicator of the future.Customer-centric, data-driven business models make modeling at the customer level, as well as understanding how best to interact with them, crucial to guiding the business.Strategic questions are about future trends in aggregate key performance indicators under different scenarios. Agent based modeling Young, growing companies

Illustrative Examples


Let’s start with a classic example that shows complex macro-level behavior emerging from a simple set of agents and some basic interactions between them. As we will discuss more later, since the focus is on helping users understand the system dynamics, we put a lot of weight on trying to make the user interaction simple, enjoyable and enlightening.

In this simulation, each circle represents an agent with a location and a direction. The agents are initialized randomly. At each timestep, each agent slightly modifies its direction according to the directions of those around it, according to the algorithm below.

speed influence constant influence distance exponent influence max angle flock by color T F number agents number colors run setup

(Note that the environment wraps around, so circles disappearing on one side reappear on the other.)

# direction modification at each timestep (translated to Python for readability)
def update_direction (agent, other_agents):
    for o in other_agents:
        if (agent['id'] != o['id']) and (agent['color'] == o['color']):
            dist = ((agent['x'] - o['x'])**2) + (agent['y'] - o['y'])**2))**0.5
            weight = 0.00005 / (dist**2)
            delta_angle = o['angle'] - agent['angle']
            agent['angle'] += max(-5, min(5, delta_angle * weight))


This next example highlights the ability of agent based models to show links between significant macro-level patterns and the often subtle preferences that guide micro-level behaviors of individuals. The nature of these links may not be intuitive; interactive simulation helps to clarify our understanding of the system dynamics. This example also provides a good opportunity to discuss the coupling of interactive simulations with the results of offline analysis with the same model.

The following simulation is an implementation of Schelling’s segregation model from 1971, which was originally carried out manually with a checkerboard. This model illustrates how urban segregation can happen even with fairly limited preference for people to live beside similar people.

speed threshold time % agents moving each timestep average % neighbors same color number agents percent lots open run setup
# decision made by each agent at each timestep
def to_stay_or_move (agent, direct_neighbors, open_spaces):
    n = sum([ d['color'] == agent['color'] for d in direct_neighbors])
    ratio_neighbours_same_as_me = n / float(len(direct_neighbors))
    if ratio_neighbors_same_as_me < threshold:
        agent['location'] = random.sample(open_spaces,0)[0]

As a complement to these interactive simulations, we can also provide decision makers with outputs more akin to the traditional analytical outputs. The models and their outputs can be made subject to many of the same interrogation techniques applied to other models used in data science, including Monte Carlo analysis and various sorts of sensitivity analysis and optimization methods.

threshold number agents percent lots open timesteps % agents moving each timestep average % neighbors same color The simulations above involverandom starting points and various random elements. The interactive graph to the right shows the summary results of Monte Carlo runs through the simulationsabove. 100% 25% 75% 50% 0% 5 10 20 15 0 You may have noticed - either by playing with the simulationor through the Monte Carlo results - that once the thresholdgets above a certain level, the ratio of neighbors with the same color begins to decline and the agents move more frequently.This graph shows the results at timestep 50 for the case where there are 100 agents and 15% probability that a lot is open. threshold 30% 50% 90% 70% 10% 100% 25% 75% 50% 0% 10% 90% 5% 25% 25 625 mean extremes middle 68% of sims middle 95% of sims

In addition to the various interrogation techniques that can be used with the models, agent based modeling also overlaps with other existing data science workflows in that agent based models can (and often should) use very detailed boundary conditions, initial state values and parameter values derived from (or inferred from) historical data.

As such, we feel that agent based modeling fits within our remit as data scientists. But how does a data scientist (or data science group) get started with it and integrate it into their toolkit?

Existing Tools

There is a plethora of existing tools for agent based modeling. Among the most popular are NetLogo, Repast, MASON and AnyLogic, the first three of which are free. Depending on your modeling needs, one of these tools may be suitable. You may also want to consider tools for “multi-agent modeling” from other fields such as robotics (e.g., Ptolemy II).

But for the data science community, it would be nice to have something in R or Python so that we could more easily integrate it with our existing workflows. For this post we will focus on Python based libraries as I’m biased towards Python :). The mesa project is worth considering (particularly if you have already made the switch to Python 3). You may also want to look at the more general simulation library SimPy.

Building in Python

Many large agent based models are custom-built to suit the problem and platform, without the use of existing tools or libraries. And even if you do choose one of the existing tools above for your project(s), it is still a worthwhile exercise to build some models from scratch in Python. The concept of agent based modeling maps very well to object-oriented programming: types of agents are classes, individual agents are objects which can have states (object variables) and response methods (object methods).

You can find a detailed tutorial on using Python to build a Schelling segregation model here. Below we present a more minimal agent based model for demonstration.

# agent definitions

class PingPongAgent():

    def __init__(self, sound):
        self.sound = sound

    def respond(self,msg):
        if msg == 'hit_it':
            return self.sound
        return None

# model definition

class PingPongModel():

    def __init__(self):
        self.current_turn = 0
        self.agents = []
        for agent_sound in ['ping','pong']:

    def simulate_timestep(self):
        output = self.agents[self.current_turn].respond('hit_it')
        self.current_turn += 1
        if self.current_turn >= len(self.agents):
            self.current_turn = 0
        return output

>>> model1 = PingPongModel()
>>> for i in range(5):
...     print model1.simulate_timestep()

A further refinement on this simple example is to use the multiprocessing package from the standard Python library. This is an essential step in moving toward distributed computing for models with hundreds, thousands or even millions of agents.

import multiprocessing
import random

# agent definitions

class PingPongAgent(multiprocessing.Process):

    def __init__(self, msg_pipe, sound):
        super(PingPongAgent, self).__init__()
        self.msg_pipe = msg_pipe
        self.sound = sound

    def run(self):
        running = True
        while running:
            msg = self.msg_pipe.recv()
            if msg == 'stop':
                running = False
            if msg == 'hit_it':

# model definition

class PingPongModel():

    def __init__(self):
        self.current_turn = 0
        self.agents = []
        self.msg_pipes_to_agents = []
        for agent_sound in ['ping','pong','punt','pass','play']:
            parent_conn, child_conn = multiprocessing.Pipe()
            p = PingPongAgent(child_conn, agent_sound)

    def simulate_timestep(self):
        this_turn = random.sample(range(len(self.agents)), 
        for i in this_turn:
        output = []
        for i in this_turn:
            output.append( self.msg_pipes_to_agents[i].recv() )
        return output

    def terminate_simulation(self):
        for i,a in enumerate(self.agents):

>>> model1 = PingPongModel()
>>> for i in range(5):
...     print model1.simulate_timestep()
['pass', 'pong', 'ping', 'play']
['pass', 'pong']
['punt', 'pass']
['pong', 'pass']
['pong', 'pass', 'punt']
>>> model1.terminate_simulation()

Watch out for the overhead imposed by process proliferation: you will likely want to group agents within processes as the number of agents increases. Further challenges await if you move to a distributed environment. And for speed, you may also want to rewrite ssome core components in a lower-level compiled language, use numba, or break from agent based orthodoxy in some places by using numpy arrays instead of sets of agent objects. But you can probably see a general path from here, or at least by noodling through some code along these lines you will be better informed about as you choose from the tools noted above.

Browser-Based Interfaces

The goal is to improve decision makers’ understanding of the dynamics of a problem. Easily accessible, beautiful and intuitive model interfaces go a long way in this direction.

The NetLogo interface has been a major part of that tool’s popularity. (The interactive simulation interfaces shown above are heavily influenced by it.) Laying out the interface of simple sliders, toggles, buttons and real-time output displays is an essential part of constructing any NetLogo model. NetLogo Web allows you to play with NetLogo models directly in the browser. So if you are willing to learn the Logo language, find the interface suitable, and your models are not too large, then this may be a good complete solution for you.

Other off-the-shelf options for browser-based interfaces include AgentScript, which implements NetLogo’s semantics in coffeescript / javascript, and AgentBase, which is also in coffeescript and is also inspired by NetLogo but breaks with its semantics. Or, of course, you can just write your models directly in javascript with some D3.js, as done for the simple examples above. (Entertaining, but maybe not for everyone.) But note that these all involve running the model directly in the browser, which limits the model’s scale.

However, a large-scale solution that many data scientists will find familiar is to write the interface in javascript / D3 and the model itself in Python, and to connect the javascript front-end via an API to the Python model running on a server (or collection of servers). In most ways this is similar to setting up a data visualization dashboard, except for the management of long-running processes and the treatment of state - not insignificant challenges, but both manageable.

Closing Thoughts

Agent based modeling is often overlooked as a potential tool in the data science community. But it does append itself nicely to other techniques and languages used by data scientists. Through the use of interactive simulation it can aid decision makers to better understand the problem or system under consideration, which ultimately extends the ways in which data science can help us to make better decisions.

1 Perhaps its greatest potential as an analytical modeling tool (independent of interactive simulation) is in risk assessment, as noted in this Economist article on the use of agent based modeling for financial system analysis.

Tweet this post! Post on LinkedIn

Come Work with Us!

We’re a diverse team dedicated to building great products, and we’d love your help. Do you want to build amazing products with amazing peers? Join us!