This directory archives Fast Downward experiments.

Experiments using Downward Lab
==============================

All experiments use their own common_setup.py module that simplifies the
creation of experiments. We recommend using common_setup.py (and related
*.py files) from issue899 for new experiments since it is the most
recent version.

Below we list some experiments that show how certain tasks can be
performed. Some of them use Downward Lab's deprecated
DownwardExperiment:

- Test changes that affect the whole planner:
  - issue481/v1-*.py show the general setup. You will need different
    experiments for satisficing, optimal and anytime configurations.

- Add a custom log parser:
  - issue735/v1.py

- Compare all attributes for tasks where we lose coverage:
  - issue439/regressions.py

- Use custom time limit:
  - issue439/issue439.py

- Add scatter plots for custom attributes:
  - issue214/issue214.py

- Run configurations on the same tasks multiple times to reduce noise:
  - issue420/issue420-v1-regressions.py

- Independent CompareRevisionReports for portfolio configs and core solvers:
  - issue462/issue462-opt.py

- RelativeScatterPlotReport:
  - issue77/issue77-v7-opt.py

- M&S: configs and additional parsing and a check for availability of matplotlib
  - issue849/v1.py and issue849/ms-parser.py

- Run experiment in debug mode (with assertions):
  - issue650/v2.py

- Translator:
  - issue862/v5.py is a translator-only experiment, including a report
    that only shows results where the translator output changed
  - issue887/v1.py is essentially the same, but with a before/after
    comparison rather than a "plain" multiple-configuration report.
  - issue862/v5-planner.py is a follow-up experiment to v5.py to see
    the impact of the translator changes on the overall planner; it is
    essentially a standard planner experiment though, not very
    translator-specific; the subset of domains to evaluate is
    specified manually


Microbenchmarks
===============

Some experiments don't run the whole planner or planner components, but
just contain small microbenchmarks for particular functionality. These
may be a good starting point for similar microbenchmarks. Examples:

- Benchmarking of random number generation:
  - issue269/rng-microbenchmark

If you add your own microbenchmark, it is recommended to start from a
copy of an existing example and follow the naming convention
issue[...]/[...]-microbenchmark for the code. This way, .hgignore
should be set up correctly out of the box.
