Skip to content
This repository

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP

GAIN Artificial Intelligence Network

branch: master

This branch is 0 commits ahead and 0 commits behind master

Fetching latest commit…

Cannot retrieve the latest commit at this time

README.md

GAIN - Artificial Intelligence Network

GAIN is a project in progress. The goal is to create a tool akin to Apache Mahout, but written around Erlang rather than Java. The initial focus is to understand and implement whatever algorithm there is in Mahout and then later look at performance, speed, and so on.

Starting out

This is a fast startup description written by jlouis so we can get people up to speed rather quickly.

Getting and running riak.

We have initially started with the riak k/v store as the data backend. Riak is available from github at this link. To install riak, just follow their basic instructions which will create an embedded release in the rel subdirectory. At the moment we simply then start a riak node at riak@127.0.0.1, like their hints suggest.

There is one specific configuration parameter that has to be set at the moment on riak. This is the following:

  {vnode_cache_entries, 0},

which disables the riak-internal LRU cache completely. We have seen a number of problems with this internal caching system, so we recommend it to be turned off at the time being.

Getting the reddit raw data.

For now, see the README.md file in the reddit/ subdirectory. Then run 'make' in that directory to generate a simplified vote database we use to load the data into riak. Note we are calling sort with a memory usage of 400 megabytes, so you need some free memory to do the sorting step.

Loading data into riak.

First, we assume you already have a riak instance up and running. And that you have tested it for the basics of storing data and retrieving data from it.

Then you run

 make deps # Fetch dependencies
 make rel

which creates a release of gain in the rel subdirectory. Now you can run

 rel/gain/erts-<vsn>/bin/erl -pa apps/gain/ebin

giving you an erlang containing the gain application as well as the riakc protocol buffers client. The '-pa' addition of the ebin path is not strictly necessary, but it makes it easier to 'l(Mod)' from the shell to get the latest version of a module when you compile it in the app.

At this point, try to execute the command

 reddit_rec:read_file("reddit/affinities.erl.small.dump").

in the erlang shell to dump the reddit data in a bucket <<"affinities">> inside riak. At the moment, this where we are.

Test cases

We have a target for running test cases. At the moment, we are just using eunit tests in modules when we need to do some tests. The target to run is,

 make eunit

which will run all eunit tests.

K-Means

We have a preliminary implementation of a (Voronoi-) K-means algorithm. We assume that a bucket 'B' exists. B contains {K, V} objects where K is a row name designator and V is a sparse vector stored as a list [{name(), float()}] where we name each dimension and supply a floating point value in the range -1 to 1. Note that we take as a precondition that the names are ordered according to the Erlang standard ordering operator <.

K-means then run in a step-iteration algorithm based upon the idea of Lloyd. Note that the first step is to sort each {K, V} pair into the cluster closest to the pair. The second step is to calculate a new mean for the cluster based upon its members.

The first step is a map in a Map/Reduce implementation: Simply for each pair, check all clusters and report the one it is closest to. The second step is a reduce step: Gather up elements belonging to each cluster and calculate the mean. Note that several nice properties of the mean w.r.t. commutativity make it possible to form a reduction tree.

Finally, to initialize the algorithm, produce K clusters by randomly assigning values in the N-dimenional space I^N where I is the interval [-1:1]. Iterate until the K clusters differ from the earlier iteration by a value less than some small epsilon.

This idea has been implemented in gain_kmeans.erl.

Testing

Assume the data above has been loaded from reddit into the bucket <<"affinities">>. Then we can ask for a kmeans run on these data via the command

 reddit_rec:kmeans().

which will try to run a K-means computation on the reddit data.

Something went wrong with that request. Please try again.