dynamic-context-decision-making

Mike Shvartsman, Vaibhav Srivastava, & Jon Cohen.

Library for building of models of decision making with dynamic context. If you use it, you should email me at [email protected]! Not only would it make me happy, but it would mean I can tell you when new and exciting improvements happen.

Installing:

You will need cmake and a compiler that plays nicely with c++11. On OSX I highly recommend getting a c++11-friendly GCC via http://hpc.sourceforge.net/, and cmake via homebrew (follow instructions at http://brew.sh/, then brew install cmake). Then:

mkdir bin
cd bin
cmake ..
make cddm

The cddm target will compile a shared library libcddm.dylib (or hopefully the equivalent on your platform). To use the library you want to import cddm_main.h, and add the library to your library paths. Right now everything is very fragile so watch your paths.

If your c++-11 friendly compiler is not your default then you will need to pass that into cmake -- for example, in my system:

...
CXX=g++-4.9 cmake ..
...

I also personally use ninja instead of make, because I find it faster. To do that you would install ninja (brew install ninja) and pass that to cmake with the -G flag (as in cmake -G Ninja), and then use ninja instead of make.

Unit tests:

I use the Catch unit-testing framework. To build all the tests (235 and counting!), you build the catch_main target and run it:

make catch_main
./catch_main

Building documentation:

make doc builds doxygen documentation in html and latex format. Right now latex gets confused by the out-of-source doc build so the TOC is broken, but html is the preferred way to go anyway.

Example tasks:

Some examples task implementations exist in examples/. There are also compile targets for them: flanker_batch, axcpt_batch , flanker_trace and axcpt_trace, and an examples target that builds all of them. Both tasks are those used in the NIPS submission.

The trace variants are meant for generating full random walk trajectory traces for visualization and testing: they output a big pile of CSVs. R/belief_visualizer.R has some code for visualizing AX-CPT, but it is not well documented or maintained. R/nips2015_plots.R has code for generating the NIPS simulation figures, assuming you have built the examples target. Simplest way to run it is to navigate to the R/ directory and call Rscript nips2015plots.R.

The event variants have a little less detail: no full trajectories, but everything else that is in trace: individual RTs and accuracies for each trial, as well as other things like the time when the inference started/stopped on each trial.

The batch variants are meant to be fast for cluster execution and work as listeners: they wait for a row of input parameters, output results, and then wait for another row of parameters. A newline exits, and # is a comment character (mostly implemented for undocumented config file functionality). Parameter format is key=value,key=value. All the parameters supported are under the relevant *_runner.cpp files under examples. For one-and-done usage you can go echo "par1=val1,par2=val2" | ./axcpt_batch or just echo "#" | ./axcpt_batch to run with defaults.

The interfaces are all a bit hackish, sorry.

Useful parameters to pass into the examples:

• decisionThreshold should sanely be between 0.5 and 1 (not inclusive), and defines the threshold in posterior probability space for the random walk. Realistic values are in the .9s.
• contextNoise is a scalar, isomorphic to sample rate on context sampling. Values in the 2-5 range give reasonable RTs with actual sample rate fixed at 10ms.
• targetNoise is a scalar, isomorphic to sample rate on target sampling. Values in the 2-5 range give reasonable RTs with actual sample rate fixed at 10ms.
• maxTrials is the number of trials to run. 1000 is enough to exclude bad points, 5000-10000 or more is more reasonable to tease apart things at the top. The reason it is maxTrials and not just nTrials is because there are eventual plans to enable early stopping based on monitoring the current estimate variance. Not implemented yet, though.