Soar 9.3.2 Release Notes, April 2012
This release of Soar continues the 9.3 line which includes modules for
reinforcement learning (RL), episodic memory (EpMem), and semantic
memory (SMem), as well as everything from previous versions of Soar. It
includes a new "forgetting" mechanism, several bug fixes, a new distribution
structure, and a simplification of how libraries are built in Soar. All
learning mechanisms are disabled by default.
Soar can be downloaded from the Soar home page at:
Soar releases include source code, demo programs, and a number of
applications that serve as examples of how to interface Soar to an
external environment. There is support for integrating Soar with C++,
Java, and Python applications. Many tools that aid in development of
Soar programs are also available. The newly expanded download section
of the wiki allows you to browse and download all of the different
distributions, tools and agents.
[Help and Contact information]
Please do not hesitate to contact the Soar group by sending mail to:
For general Soar-related discussion and announcements:
For more technical developer discussion:
Please do not hesitate to file bugs on our issue tracker (search for
Important Changes for 9.3.2
Soar's build process has undergone significant modifications, and as a result
several things have changed in the binary release:
1. Instead of separate libraries for ClientSML, ConnectionSML, ElementXML,
KernelSML, and the Soar Kernel, there is now only a single Soar library. On
Windows this is Soar.dll, on Linux libSoar.so, and on OSX libSoar.dylib. This
means that when you are building SML clients, you only need to link to that
single library. The SML header files have not changed: you still need to include
2. Due to the library merge, SML clients no longer have to locate the KernelSML
library at run time. Therefore, all references to the location and name of that
library, such as DefaultLibraryLocation, have been removed from the SML API.
- The set-library-location CLI command has been removed
- The following functions have been removed:
- The library location argument has been removed from the signatures of the
sml::Agent::SpawnDebugger (see below)
3. The sml::Agent::SpawnDebugger function no longer uses the KernelSML library
location to find SoarJavaDebugger.jar. Instead,
1. If the path of the jar (not of the containing directory) is provided as
an explicit argument, it is used.
2. Otherwise, if SoarJavaDebugger.jar is in the current working directory,
then that jar is used.
2. Otherwise, if the environment variable SOAR_HOME is defined and points
to the directory that contains SoarJavaDebugger.jar, then that jar is used.
3. Otherwise the search fails and the function returns false.
4. Due to the reduction in the number of files, we have merged the lib directory
into the bin directory. The Soar library and the SWIG wrapper libraries now
reside in the same directory as the executables. This makes life easier for
Windows users running programs such as TestCLI or the Java debugger, because
Windows searches the current directory for library dependencies by default.
5. Linux and OSX users no longer need to set LD_LIBRARY_PATH and
DYLD_LIBRARY_PATH to point to the bin directory when running the native
executables in the distribution, such as TestCLI or TestSoarPerformance. The
relative location of the library has been hard-coded into those executables
using the rpath and install_name mechanisms in Linux and OSX, respectively.
However, those environment variables must still be set for running the java
debugger, as required by the JVM.
6. SoarSuite is now extremely streamlined. It is intended for users
who want to either (1) just run Soar or (2) plug-in the latest version
of Soar into their current work. It only contains the Soar kernel, the
Soar Java Debugger, Visual Soar, the Soar Manual, the default rules,
unit test programs and the Soar header files.
7. The Soar Tutorial now has its own download. It contains the Soar kernel,
the Java Debugger, Eaters, TankSoar, the Soar Manual, the Soar Tutorial,
VisualSoar, and all the introductory agents.
8. Nearly everything we provide now has stand-alone versions, including TankSoar
and Eaters. All agents, tools and environments now have their own wiki pages
with thorough descriptions and individual downloads. We highly encourage
people to make their own work available on the Soar wiki.
9. The Soar SVN repository has went through major restructuring and clean up
to support these changes.
10. When a tie impasse occurs, the substate is augmented with two new WME
classes, which document operators that do not have numeric-indifferent
preferences associated with them. These augmentations are
"non-numeric" and "non-numeric-count" and operate analogously to the
existing "item" and "item-count" augmentations.
11. We have added a new forgetting mechanism to Soar to automatically
excise chunks via the base-level decay model. It is controlled via the
"rl" command using the "apoptosis" parameters. This functionality
helps develop learning agents that can persist for long periods of
time, accumulating a large amount of procedural knowledge while
bounding memory consumption. Version 9.3.1 of Soar included forgetting
of WMEs via the "wma" command, which served to bound procedural
matching and episodic reconstruction. WME forgetting worked well with
automatic semantic encoding via the "mirroring" parameter of the
"smem" command. Together, these forgetting policies facilitate the
development of long-living, learning agents that remain reactive in
A more comprehensive list of changes is below.
The scons build process now works in Windows and is the preferred way to build
Soar on all three supported platforms.
The old Visual Studio project and solution files have been removed from the
distribution. Scons can generate those files to make it easier for Windows users
to debug the kernel. Scons is now also included in all source distributions to
simplify building Soar.
Ant is no longer needed to build the Java debugger.
Soar-specific environment variables are no longer needed to compile Soar on
Linux and OSX.
Improved documentation and completely restructured Soar web page and wiki.
A lot of outdated information has been updated or removed.
Demo agents have been removed from the Java debugger and SoarSuite. They are now
simply packaged in the Agents directory of the SoarTutorial. They are also
available for individual download.
New chapters have been added to the tutorial for episodic and semantic memory.
VisualSoar has been removed from Soar core but can still be downloaded from its
own wiki page.
SoarSuite now only includes the Soar kernel, the Soar Java Debugger, Visual
Soar, the Soar Manual, the default rules, unit test programs and the Soar
Added a new download specifically for the SoarTutorial. It includes the
Soar kernel, the Java Debugger, Eaters, TankSoar, the Soar Manual, the
Soar Tutorial documents, VisualSoar, and all the introductory agents.
There are now stand-alone versions of TankSoar and Eaters. They include
self-contained copies of the Soar kernel and the debugger and can be run from
A large variety of agents are now available for individual download. They
each have their own wiki page with a complete description.
Default rules updated for Soar 9.
Soar now works with the OpenJava JDK, which is what is installed by default on
The debugger no longer has a Demos menu and has had several demo-related buttons
All releases for the various platforms now have identical naming and directory
structures. This should make things much more consistent and less brittle.
Issue 78: The selection probabilities outputted by the "pref" command now
make sense for all exploration strategies, not just softmax.
Issue 92: Removed the "run forever" command. "run" without any arguments will
achieve teh same thing.
Issue 93: Alias command now adheres to the documentation. First argument is
once again the alias name.
Issue 95: Certain types of comments on the RHS of productions no longer
produces an error in the parser.
Issue 98: Soar will no longer automatically change the indifference-selection
policy to episilon-greedy when turning reinforcement learning on. This may
break the expected behavior of some agents, so users should make sure to set
the policy the want explicitly. All of the agents in our repository have been
updated to explicitly set it to epsilon-greedy after they turn RL on.
Issue 98: To increase performance, Soar will no longer discount updates over
subgoals as though they were gaps, i.e. HRL-Discount is now off by default.
Issue 98: By default, Soar will no longer create a chunk that only differs
from another rule in the value of the numeric indifferent preference, i.e.
Chunk-Stop is now on by default.
Issue 102: CLI now strips out comments at the end of the line properly.
Issue 103: Watch-wmes command now works again. You can now set up arbitrary
filters on which wmes are printed out.