Focused Meta Actions

This is the code repository of the master thesis Focused Meta Action by Kristian Skov Johansen and Jan M.D. Hansen. This readme will serve as a way to navigate and understand the different sections of the project. Generally, this project is split into a few couple of repositories, where the most important ones are:

• This one
  • The Training folder contains file and project needed for training.
  • The Testing folder contains file and project needed for testing.
    • This folder also contains log files from the learning process, as well as the final meta action sets.
  • The Evaluation folder contains all the automatic graph and table generation used in the paper.
    • This folder also contains the final CSV files that the results in the paper where generated from.
• A benchmark repository
• A repository for labyr (do note, this is not needed to run the project, its only used to run all the experiments from the paper)
• A repository for a modified Stackelberg Planner for state exploration and macro generation
• A repository containing the main reconstruction method (Do note, this is from a previous semester project and is not 100p compatible with the formats of this project)

All the C# side uses a lot of NuGet packages made for this project. NuGet should be able to install them themselfs, but to be on the safe side, here is the list of packages:

• PDDLSharp to manipulate and parse PDDL files
• CSVToolsSharp to output CSV files
• MetaActionGenerators to generate initial meta actions
• Stackelberg.MetaAction.Compiler to compile domain+problem+meta action into a Stackelberg variant.

Do note, that this entire toolchain was made to work on linux, you may be able to use it on windows (if you can get the Stackelberg Planner and Fast Downward to compile) if you need to. Otherwise, we recommend windows users to set up a WSL instance to run it on.

Setup

Firstly to run this entire toolchain, do cd ./Dependencies and execute the setup.sh script. This should install all needed dependencies for the project.

When all the dependencies are installed, you can compile the training program by writing:

dotnet build --configuration Release --project Training/FocusedMetaActions.Train/FocusedMetaActions.Train.csproj

It should pop up with a bunch of different arguments you should give it before it can run. Each argument have a little description that should help you.

Examples for Training

Here are a few examples of arguments you could give to the training to run on (this is run from the Training folder):

dotnet run --configuration Release -- --output output --domain ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/domain.pddl --problems ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/training/p1.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/training/p2.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/training/p3.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/training/p4.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/training/p5.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/usefulness/p1.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/usefulness/p2.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/usefulness/p3.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/usefulness/p4.pddl ../Dependencies/focused-meta-actions-benchmarks/Benchmarks/blocksworld/usefulness/p5.pddl --generator CPDDLMutexed --args cpddlOutput;../Dependencies/focused-meta-actions-benchmarks/CPDDLGroups/blocksworld.txt --refinement-time-limit 120 --exploration-time-limit 120 --validation-time-limit 20 --cache-generation-time-limit 20 --pre-usefulness-strategy UsedInPlans --post-usefulness-strategy ReducesMetaSearchTimeTop2 --last-n-usefulness 5

Examples for Testing

For testing, you can use the labyr package. This is basically a small package made to run training+testing with and compare against Fast Downward. This package uses script files to run. There are two script files in this repository, designed for the paper benchmarks. You can run the script from the root of the project as follows:

dotnet build --configuration Release
labyr run-experiments-full.toml --work-dir Results --threads 0 --keep-working-dir

This will create a results folder in the end, containing the combined results of all the runs.

How to run this on the Cluster

The only difference is in test execution. Namely, the execution method labyr should be slurm, i.e. give the argument -e slurm.

dotnet build --configuration Release
labyr run-experiments-full.toml -e slurm --work-dir Results --threads 0 --keep-working-dir