OpenCog is a framework for developing AI systems, especially appropriate for integrative multi-algorithm systems, and artificial general intelligence systems. Though much work remains to be done, it currently contains a functional core framework, and a number of cognitive agents at varying levels of completion, some already displaying interesting and useful functionalities alone and in combination.
The main project site is at http://opencog.org
OpenCog consists of multiple components. At its core is a (hyper-)graph database, the AtomSpace, which is used for representing knowledge and algorithms, providing a surface on which learning and reasoning algorithms are implemented. The AtomSpace consists of an in-RAM database, a "query language" aka "pattern matcher", a (ProLog-like) rule system, including forward and backward chainers, and an evaluator for the internal "programming langauge", Atomese. This language is not really meant to be used by humans (although, defacto, it is) but rather, it is a language for representing knowledge and algorithms, on which (automated) reasoning and learning can be performed. The AtomSpace also provides Scheme (guile) and Python bindings. The AtomSpace is maintained in a separate git repo: http://github.com/opencog/atomspace
This git repository contains assorted projects that are central to the OpenCog project, but are not yet mature or stable, and are subject to active development and experimentation. These include:
- An assortment of natural language processing subsystems, including: -- Natural language generation (for expressiong thoughts as sentences). -- Natural language input (for reading and hearing). -- Assorted chatbots, some of which are embodied.
- PLN, a probabilistic reasoning and inference system.
- Attention Allocation, for managing combinatoric explosion during reasoning and language generation.
- Space-time servers, for managing spatial and time data (grounding common-sense natural language concepts such as "next-to", "nearby", and "soon".)
- An embodiment subsystem, attaching language to visual and auditory senses. This is primarily located in the ROS Behavior Scripting repository.
- OpenPsi, a model of psychological states. Its currently a mashup of two unrelated ideas: a generic rule-class selection and plannning system, and a model of human psychological states. An open to-do item is to untangle these two.
- An unsupervised learning system or "pattern miner", for extracting "surprising" patterns.
- A supervised learning system, MOSES, for extracting patterns from tabular data. This is located in a seprate repository, MOSES.
- The CogServer, a network server providing shell access and a REST API.
- Several (obsolete!?) data visualization subsystems.
With the exception of MOSES and the CogServer, all of the above are in active development, are half-baked, poorly documented, mis-designed, subject to experimentation, and generally in need of love an attention. This is where experimentation and integration are taking place, and, like any laboratory, things are a bit fluid and chaotic.
Building and Running
For platform dependent instruction on dependencies and building the code, as well as other options for setting up development environments, more details are found on the Building Opencog wiki.
There is no single "demo" or system that can be "run"; rather, the
various subsystems can be run individually, or together. The single
most-fully-integrated, complete demo would be the embodied Hanson
Robotics chat subsystem. This
can be run without having an actual robot; a virtual Blender
animation may be used instead; a webcam and microphones are required
for sensory input. Portions of this system can be found in the
directory, in this repo, as well as the
ROS Behavior Scripting
repo. The full setup is located in the Hanson Robotics
HEAD repo, and ready-to-run
Docker images can be found in the OpenCog Docker
To build and run OpenCog, the packages listed below are required.
With a few exceptions, most Linux distributions will provide these
packages. Users of Ubuntu 14.04 "Trusty Tahr" may use the dependency
/scripts/octool. Users of any version of Linux may
use the Dockerfile to quickly build a container in which OpenCog will
be built and run.
Common OpenCog C++ utilities http://github.com/opencog/cogutil It uses exactly the same build procedure as this package. Be sure to
sudo make installat the end.
OpenCog Atomspace database and reasoning engine http://github.com/opencog/atomspace It uses exactly the same build procedure as this package. Be sure to
sudo make installat the end.
The following packages are optional. If they are not installed, some optional parts of OpenCog will not be built. The CMake command, during the build, will be more precise as to which parts will not be built.
Natural Language Parser for English, Russian, other languages. Required for natural language generation, and the chatbot. http://www.abisource.com/projects/link-grammar/
MOSES Machine Learning http://github.com/opencog/moses It uses exactly the same build proceedure as this package. Be sure to
sudo make installat the end.
3D occupancy grid mapping library Required for the robot perception subsystem.
sudo apt-get install liboctomap-dev
The following packages are needed to build some of the old, obsolete packages.
C++ HTTP RESTful interfaces Used by the Pattern miner for distributed processing (this will be replaced by gearman in future releases).
sudo apt-get install libcpprest-dev
Threading Building Blocks
C++ template library for parallel programming Used to implement the optional REST API. (TODO: the REST API should be refactored to not use TBB)
sudo apt-get install libtbb-dev
Perform the following steps at the shell prompt:
cd to project root dir mkdir build cd build cmake .. make
Libraries will be built into subdirectories within build, mirroring the structure of the source directory root.
To build and run the unit tests, from the
./build directory enter
(after building opencog as above):
Some useful CMake's web sites/pages:
- http://www.cmake.org (main page)
The main CMakeLists.txt currently sets -DNDEBUG. This disables Boost matrix/vector debugging code and safety checks, with the benefit of making it much faster. Boost sparse matrixes and (dense) vectors are currently used by ECAN's ImportanceDiffusionAgent. If you use Boost ublas in other code, it may be a good idea to at least temporarily unset NDEBUG. Also if the Boost assert.h is used it will be necessary to unset NDEBUG. Boost ublas is intended to respond to a specific BOOST_UBLAS_NDEBUG, however this is not available as of the current Ubuntu standard version (1.34).
-Wno-deprecated is currently enabled by default to avoid a number of warnings regarding hash_map being deprecated (because the alternative is still experimental!)