Implement a relevant subset of precondition choice functions from https://fai.cs.uni-saarland.de/lauer/papers/hsdip2020.pdf . (Likely just GZD+BD.)

Pull request:

https://github.com/aibasel/downward/pull/301

I attached the reports with the overall results and pairwise diffs between all configurations.

The new results show a good increase in coverage overall. I could not find anything particularly suspicious in the diffs.

We merged the changes in Commit: https://github.com/aibasel/downward/commit/e2f4adc290814c20e9e41539d539fd5ba5d6d404

There are some open questions about if we could improve the code further. I will create separate issues for that.

And now also the version on infai_2 is done:

https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/issue1228-exp2-eval/issue1228-exp2.html
https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/

For the new experiment, the version that was run on infai_3 is done:

https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/issue1228-exp2_infai_3-eval/issue1228-exp2_infai_3.html
https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/

Note that the previous experiment was run on different hardware, infai_2. The new experiment is still running on that hardware, around 80% done at the moment.

I reviewed the pull request today. Apart from one function name which I find unclear, everything looks good to merge.

First experimental results are in:

https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/issue1228-exp1-eval/issue1228-exp1.html
https://ai.dmi.unibas.ch/_experiments/ai/downward/issue1228/data/issue1228-exp1-eval/

This form of reporting results is a bit annoying with all the configs that failed because of unsupported arguments, and we should probably produce two separate tables from this data, one that compares the old default version before/after and one that compares the four configs in the new code.

I would then also like to see some relative comparison tables that give a better picture of what happens per domain. At first glance it looks like there is some performance loss for the old default config, and perhaps we can look at that a bit more closely.

But overall the coverage numbers look good.