Skip to content
Human-Computation Runtime
Scala Java
Find file

Circle CI

AutoMan: Human-Computation Runtime v1.0

This is a major release of AutoMan. Code written using earlier versions (< 1.0) will need to be rewritten as AutoMan's syntax has changed. See below for details.

What is AutoMan?

AutoMan is the first fully automatic crowdprogramming system. AutoMan integrates human-based ("crowdsourced") computations into a standard programming language as ordinary function calls that can be intermixed freely with traditional functions. This abstraction lets programmers focus on their programming logic. An AutoMan program specifies a confidence level for the overall computation and a budget. The AutoMan runtime system then transparently manages all details necessary for scheduling, pricing, and quality control. AutoMan automatically schedules human tasks for each computation until it achieves the desired confidence level; monitors, reprices, and restarts human tasks as necessary; and maximizes parallelism across human workers while staying under budget.

AutoMan is available as a library for Scala.

Getting AutoMan

The easiest way to get AutoMan is via the Maven Central Repository. If you're using SBT:

libraryDependencies += "edu.umass.cs" %% "automan" % "1.0.1"

or if you're using Maven:


where X.XX is 2.10 or 2.11 depending on your Scala version.

Latest Updates

Major changes include:

  • Support for non-quality-controlled question types in addition to the standard quality-controlled types.
  • Per-question budgets.
  • All Scala concurrency details are now hidden. Instead of dealing with low-level Scala Future objects, instead you pattern-match on AutoMan Outcome objects.
  • A an important but subtle bias toward accepting low-confidence answers has been eliminated by modifying AutoMan's core algorithm. The details are discussed in our upcoming CACM Research Highlights article (stay tuned!).
  • A much more capable record-and-replay engine. For example, if your program terminates abnormally, on restart AutoMan will continue as if nothing unusual happened. Earlier versions of AutoMan only memoized answered MTurk HITs, which meant that restarts would schedule new work even if HITs had been completed during the interim. The new version records the complete state of a computation on MTurk.
  • The replay engine can also be used to provide mocks for your unit tests. We will be expanding on these capabilities in the near future--stay tuned!
  • AutoMan now has a simple plugin architecture, designed to support a visual debugger currently being developed for IntelliJ IDEA largely through the efforts of our excellent Google Summer of Code intern, Bartosz Janota. This work was based on an earlier web-based prototype from a talented undergrad working in our lab, Bianca Tamaskar. If you would like to develop plugins for AutoMan, get in touch!
  • Many changes to enhance reliability.

Bug Reports.

Please report bugs using this repository's issue tracker.


AutoMan is licensed under the GPLv2, Copyright (C) 2011-2015 The University of Massachusetts, Amherst.

Using AutoMan in Your Project

In your source file, import the Mechanical Turk adapter (Scala syntax):

import edu.umass.cs.automan.adapters.mturk._

After that, initialize the AutoMan runtime with an MTurk config:

val a = MTurkAdapter { mt =>
  mt.access_key_id = "my key"
  mt.secret_access_key = "my secret"
  mt.sandbox_mode = true

and then define your tasks:

def which_one() = a.RadioButtonQuestion { q =>
  q.budget = 8.00
  q.text = "Which one of these does not belong?"
  q.options = List(
    a.Option('oscar, "Oscar the Grouch"),
    a.Option('kermit, "Kermit the Frog"),
    a.Option('spongebob, "Spongebob Squarepants"),
    a.Option('cookie, "Cookie Monster"),
    a.Option('count, "The Count")

You may then call which_one just like an ordinary function (which it is). Note that AutoMan functions immediately return an Outcome, but continue to execute asynchronously in the background.

To access return values, you must pattern-match on the Outcome, e.g.,

val outcome = which_one()

// ... do some other stuff ...

// then, when you want answers ...
val answer = outcome.answer match {
  case Answer(value, _, _) => value
  case _ => throw new Exception("Oh no!")

Other possible AbstractAnswer types are LowConfidenceAnswer if you run out of money during a computation (which gives you access to lower-confidence results), or OverBudgetAnswer in case even low-confidence answers are not possible because you didn't have enough money in your budget to begin with.

Cleanup of AutoMan Resources

Note that, due to AutoMan's design, you must inform it when to shut down, otherwise it will continue to execute indefinitely and your program will hang:


Alternately, you may wrap your program in an automan statement, and cleanup will happen automatically. This feature was inspired by the C# using statement:

    automan(a) {
      ... your program ...

We will add more documentation to this site in the near future. In the interim, please see the collection of sample programs in the apps directory.

Supported Question Types

Question Type Purpose Quality-Controlled Number of Answers Returned
RadioButtonQuestion The user is asked to choose one of n options. yes 1
CheckboxQuestion The user is asked to choose one of m of n options, where m <= n. yes 1
FreeTextQuestion The user is asked to enter a textual response, such that the response conforms to a simple pattern (a "picture clause"). yes 1
RadioButtonDistributionQuestion Same as RadioButtonQuestion. no user-defined
CheckboxDistributionQuestion Same as CheckboxQuestion. no user-defined
FreeTextDistributionQuestion Same as FreeTextQuestion. no user-defined

Using AutoMan with a Different Crowdsourcing Backend

We currently only support Amazon's Mechanical Turk. However, AutoMan was designed to accommodate arbitrary backends. If you are interested in seeing your crowdsourcing platform supported, please contact us.


AutoMan saves all intermediate human-computed results by default. You may turn this feature off by setting logging = LogConfig.NO_LOGGING in your AutoMan config. You may also set the location of the database with database_path = "/path/to/your/database". Note that the format of the database has changed from earlier versions of AutoMan from Apache Derby to H2.

Building the JAR Yourself

This release incorporates an SBT build script that can build the AutoMan JAR for you, including downloading all of AutoMan's dependencies. The build script can also build the sample applications that are located in the apps directory. These applications are the ones used in our paper.

You can build the AutoMan JAR using the following commands:

cd libautoman
sbt pack

The AutoMan JAR plus all of its dependencies will then be found in the libautoman/target/pack/lib/ folder.

Sample Applications

Sample applications can be found in the apps directory. Apps can also be built using pack. E.g.,

cd apps/simple_program
sbt pack

Unix/DOS shell scripts for running the programs can then be found in apps/[the app]/target/pack/bin/.

More Information

More detailed information is available in the following paper, published at OOPSLA 2012, included in the repo as AutoMan-OOPSLA2012.pdf.

AutoMan: A Platform for Integrating Human-Based and Digital Computation Daniel W. Barowy, Charlie Curtsinger, Emery D. Berger, and Andrew McGregor

The full citation is given below:

 author = {Barowy, Daniel W. and Curtsinger, Charlie and Berger, Emery D. and McGregor, Andrew},
 title = {{AutoMan}: a platform for integrating human-based and digital computation},
 booktitle = {Proceedings of the ACM International Conference on Object-Oriented Programming Systems Languages and Applications},
 series = {OOPSLA '12},
 year = {2012},
 isbn = {978-1-4503-1561-6},
 location = {Tucson, Arizona, USA},
 pages = {639--654},
 numpages = {16},
 url = {},
 doi = {10.1145/2384616.2384663},
 acmid = {2384663},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {crowdsourcing, programming languages, quality control},

Contact information:

Dan Barowy, Emery Berger,

Change Log

Version Notes
1.0.1 Maintenance release.
Bugfix for timeout policy calculation.
Renamed misleading distribution answer type name.
1.0.0 Major release.
Syntax changes.
New question types.
Per-question budgets.
Bias due to multiple comparisons eliminated with Bonferroni correction.
New memo engine that allows MTurk computation to be resumed without additional cost (timeouts notwithstanding).
Support for mocks.
Plugin architecture.
Numerous other changes for better reliability.
0.4.0 Maintenance release.
Switch to SBT build system. Updates for Scala 2.10.
0.3.0 Maintenance release
Buildr Buildfile, including reorganization of project directory.
0.2.1 Maintenance release
Update to work with latest MTurk API (1.6.0).
Better log output, including scheduler object logging.
New Automatic Number Plate Recognition (ANPR) app.
0.2.0 Major rewrite to simplify syntax.
0.1.0 First release.


This material is based on work supported by National Science Foundation Grant Nos. CCF-1144520 and CCF-0953754 and DARPA Award N10AP2026.

Something went wrong with that request. Please try again.