Readme

This file will attempt to detail the assumptions and workflow of the project. There is a ticket system to keep track of what has been done and what still needs to be done.

Installation

GHC

We develop in a bare Haskell Platform environment. Stack is not used at the moment, due to the complexity of importing local packages that are not yet in Hackage.

At the moment, the code needs to work with our Ubuntu 16.04 LTS (Xenial) server, which uses GHC 7.10.3. On that distribution, it should be enough to do:

sudo apt install haskell-platform{,-doc,-prof}

On other OSes, the easiest way to get this specific version is perhaps to use the generic installer:

wget -O /tmp/hp.tar.gz \
    https://www.haskell.org/platform/download/7.10.3/haskell-platform-7.10.3-unknown-posix-x86_64.tar.gz
tar xf /tmp/hp.tar.gz
sudo ./install-haskell-platform.sh

# We also need to change some flags
sed -i 's/\(.*"C compiler flags",\s*"\)\(.*\)/\1-fno-PIE \2/g
       ;s/\(.*"C compiler link flags",\s*"\)\(.*\)/\1-no-pie \2/g
       ;s/\(.*"ld flags",\s*"\)\(.*\)/\1-no-pie \2/g' \
       /usr/local/haskell/ghc-7.10.3-x86_64/lib/ghc-7.10.3/settings

Database

Since the database is SQLite3, we need the SQLite binary and libraries. On Debian-based distributions, this amounts to:

sudo apt install sqlite3 libsqlite3-dev

On Windows, you can get the required executables and DLLs at sqlite.org.

The initial live database can later be built with the database-builder.exe binary, like so:

./database-builder.exe -o advise-me.db

Web server

To run the binary locally, you can use any web server with CGI support. We can do the following to use Apache to serve CGI scripts from the /usr/lib/cgi-bin directory on Debian-based distributions:

sudo apt install apache2
sudo a2enmod cgid

For other OSes, check this guide.

Haskell environment

The source code of the project is contained in Git and Subversion repositories. To obtain it:

git clone \
    https://github.com/ideas-edu/ideas
cd ideas; make src/Ideas/Main/Revision.hs; cd -

svn checkout \
    https://ideastest.science.uu.nl/svn/ideas/Tutors/math-types

svn checkout \
    https://ideastest.science.uu.nl/svn/ideas/Tutors/Advise-Me/trunk

Install the sandbox:

cd trunk
cabal sandbox init
cabal sandbox add-source ../ideas
cabal sandbox add-source ../math-types
cabal install \
    --only-dependencies \
    --enable-tests \
    --enable-executable-profiling \
    --enable-library-profiling
cabal configure \
    --enable-tests \
    --enable-executable-profiling \
    --enable-coverage

We use make, because there are many different files and interdependencies. Reading the Makefile should give an idea of the workflow. It is also recommended to make a config.mk file, overriding the variables in the Makefile so that they point to the correct directories:

tee config.mk << EOF
IDEAS_DIR = ../ideas/src
MATHTYPES_DIR = ../math-types/src
CGI_BIN = /usr/lib/cgi-bin
EOF

Bayesian networks

To create the Bayesian networks, Genie is used. We used to interface with the SMILE library for using the networks, but that is now done in Haskell itself by transforming the original .xdsl files into a Haskell interface. See network-builder.exe.

Compiling

Now, we can compile the binaries. make processing should take care of everything for us, but of course the binaries can also be created by cabal separately.

Note that there is an xlsx cabal flag that is on by default, because building the xlsx library (used for reading human assessments) is not straightforward on every machine. If you find that the xlsx library is causing issues and you do not need its functionality, do cabal configure -flags="-xlsx" before building.

Project structure

The following directories are important to know.

• app/: Haskell executables and scripts.
• src/: Haskell sources to the Advise-Me library.
• tests/: Haskell sources to the testing suite.
• test-data/: Test input requests for the testing suite and shell scripts to send test input to the server.
• hpc-*: Haskell code coverage reports as generated by the recipe in the Makefile.
• pilots/:
  • raw/: Databases, mostly untouched as they were collected during pilot or evaluation studies.
  • processed/: Databases that are created from the raw data after the fact, by processing it in various ways using database-builder.exe. The Makefile contains recipes to create these files.
  • assessments/: Excel spreadsheets that mirror the names in the processed/ directory. These spreadsheets contain evaluations by humans of the same data. They can be used to evaluate or debug the application, using report.exe, or to change or annotate the processed data. There are also documents in this directory that are non-machine readable, containing remarks of IDEAS' output by a human examiner.
  • regressions/: This directory contains .exp files that concatenates the expected output of the processed databases. This allows for a rudimentary regression test, using diff.
• networks/: Bayesian networks created in Genie, and a supporting XML file containing translations of the labels.

Apart from the main advise-me.cgi binary, there are a couple of auxiliary binaries to use:

• The advise-me.cgi binary provides the service: you provide input via a POST or GET request, and it will respond with the information you requested. There are also additional commands that can be given to make it do other things, like rerunning or reporting. Some of these are deprecated, and they aren't documented well.

• network-builder.exe builds, given an .xdsl file from networks/, the interface file necessary for running that network in our Haskell environment. Unfortunately, it cannot itself be actually built: it depends on the Advise-Me library, which itself depends on the files that it is supposed to generate! From cabal-install version 2, I believe that we could use its autogeneration facilities. For now, as a crutch, we run app/NetworkBuilder.hs as a script — see the Makefile.

• The database-builder.exe binary is a tool to create the initial database and process existing databases. It gives us the ability to reuse input data collected from a previous run and generate new output for it, as well as annotate the database with information tables. As there are many flags and switches, call it with --help for more info.

To inspect the resulting databases or to examine statistics, there are multiple options.

• advise-me-admin.cgi provides a web interface to inspect the databases and report on statistics.

• report.exe can be used offline to compare assessments from IDEAS in the database against human assessments with the humanvsmachine subcommand. It can also count how often evidence occurs with the priors subcommand. Finally, it can generate a legacy HTML page with diagnostics info, similar to the overview in advise-me-admin.cgi.

Testing

Tests that are implemented now relate exclusively to finding the evidence. Other tests are mostly non-existent, so functionality may break without warning. (For more fine-grained information on how well the evidence matches our expectations, see report.exe.)

Rudimentary regression tests can be performed with a diff, simply to check whether the output has changed since the last update. make regressions does this for you.

cabal test runs the tasty test suite with particular example requests, to check if they still find the evidence we expect. Whenever you fix a specific bug, please add a test along with the relevant request XML.

Coverage

To inspect code coverage, do cabal clean and cabal configure --enable-coverage and rebuild the binaries that you want to test. After running the binaries, .tix files will be created (that you can optionally combine with hpc sum *.tix). From the tix and mix files, you can generate a HTML coverage index or a statistics report. For example:

hpc report
--hpcdir=dist/hpc/vanilla/mix/Advise-me-0.1
--hpcdir=dist/hpc/vanilla/mix/database-builder.exe
database-builder.exe.tix

Profiling

If you have installed the libraries with --enable-library-profiling and configured cabal with --enable-library-profiling --enable-executable-profiling, then you can build a profiling version of the main CGI binary. The Makefile contains a recipe for a PDF report.