README.md

PlannerEvalToolkit

This repo makes it easy for us to compare planners.

How to use

0. Step 0: Install the package and prepare the planning tasks.

mkdir sandbox-dir
python -m venv venv
git clone https://github.com/AI-Planning/classical-domains
source venv/bin/activate
pip install .

1. Step 1: Describe your experiment parameter.

mkdir -p exp-1/planners
touch exp-1/exp-details.json

The exp-details.json defines the following parameters: - [ ] time limit - [ ] memory limit Here is a demo exp-details.json

{
    "name": "demo",
    "cfgs": {
        "timelimit" : "00:30:00",
        "memorylimit": "10GB"
    }
}

2. Step 2: Describe which planners are used inside this experiment, this is done by creating json file inside planners dir. This is an example showing using an experiment with two planners.

touch exp-1/planners/smt-seq-planner.json
touch exp-1/planners/smt-par-planner.json

Here is the content for smt-seq-planner.json:

{
    "planner-tag": "SequentialSMT-tag",
    "up-planner-name": "SMTPlanner",
    "planner-params": {
        "encoder": "EncoderSequentialSMT",
        "upper-bound": 1000,
        "search-strategy": "SMTSearch",
        "configuration": "seq",
        "run-validation": false,
        "compliation-list": [
            ["up_quantifiers_remover", "QUANTIFIERS_REMOVING"],
            ["fast-downward-reachability-grounder", "GROUNDING"]
        ]
    }
}

Here is the content for smt-par-planner.json:

{
    "planner-tag": "ParallelSMT-tag",
    "up-planner-name": "SMTPlanner",
    "planner-params": {
        "encoder": "EncoderParallelSMT",
        "upper-bound": 1000,
        "search-strategy": "SMTSearch",
        "configuration": "forall",
        "run-validation": false,
        "compliation-list": [
            ["up_quantifiers_remover", "QUANTIFIERS_REMOVING"],
            ["fast-downward-reachability-grounder", "GROUNDING"]
        ]
    }
}

3. Step 3: Generate the experiment commands, such that each command runs a planning task with a planner. This is acheived by

pypmtevalcli generate --venv-dir $(pwd)/venv --exp-details-dir $(pwd)/exp-1 --sandbox-dir $(pwd)/sandbox-dir --planning-tasks-dir $(pwd)/classical-domains

After running this command, it generates generated_cmds.sh such that each line is a command aims to solve one of the planning tasks found in --planning-tasks-dir

4. Step 4: Running the experiment commands, which can be done by using gnu-parallels or slurm.

Note that the results are stored in sandbox-dir/dump_results dir in json file format.