aibasel / machetli Goto Github PK

• method evaluate in Evaluator
• method get_successors in SuccessorGenerator

This should make the code more consistent when defining the initial state and is necessary if we want to implement evaluators with fully-implemented evaluate(state) methods.

Make Minimizer usable as command line program for simple use cases.

The implementation of the Parser class provides the function add_pattern passes on its input to the constructor of _Pattern. However, said constructor breaks upon call as it does not expect these inputs.

probably easiest by using tarski instead of doing our own stuff
related to #60

The visitor pattern should do the recursion so that it doesn't need to be re-implemented in every inherited class.

write function for PDDL visitor body that takes a visitor and a name to replace

for recovering crashed experiments

• Research available signals and ways of communication between processes
• Research MPI/Open MPI
• Idea: Single grid node that is dedicated only for search
• Idea: Write state of search to disk instead of keeping it as an active job
• Idea: monitoring process that checks whether certain files have been written
• Idea: Evaluate batches of 100 successors

do error handling in search, so we don't need to do it in every environment

Trying successor 403120/2131151 (18%) [======> ]

Instead of passing the evaluator as a Python object to the search, we want to require it as a stand-alone script and pass the file path to this script to the search. This works more cleanly in the grid environment, where the compute nodes don't have direct access to the classes defined in the main file.

The resulting constructor could be:
Run(command, input_file=None, time_limit=1800, memory_limit=None, log_output=None)

This would reduce code redundancy. The passed environment object would then control whether the grid functions are invoked.
It might also be a good idea to rename the function from first_choice_hill_climbing to search, because this is simpler and currently, first-choice hill-climbing is the only implemented search algorithm for the Minimizer.

• Abstract search state represented by a dictionary
• Successor generator
• Evaluation function
• Write search code skeleton with abstract classes
• Store task internally instead of in the form of PDDL or SAS+ files

• Error reporting should not be part of error classes. Implement like standard errors in python and use the message member.
• abort should always lead to logging.critical
• There's too much logic in the exceptions.

Currently, the add_pattern method of the parser class has an argument cmd_names, which accepts one or multiple run object names (as, e.g., defined and used here). Storing the names of the runs inside a parser object creates an unnecessary coupling between a parser and the run(s). The parser does not need to know anything about the run(s) it should be used for.

Improvement suggestion: Remove the cmd_names argument from add_pattern and add a parsers argument to the constructor of the run class. This way, each run has an internal list containing the parser(s) to be applied to its output.

disallow submitting the same job without asking because it overwrites data

A specialized option of the Run class (e.g., FastDownwardRun) could make life easier for the user when defining a FD run by providing an interface where the planner, the PDDL files, etc. could be provided as constructor arguments instead of passing the actual entire call string to the constructor.

remove functions that are no longer used

Users might want to generate successors by applying a combination of several successor generators. The Python itertools module could be helpful for this.

The installation of minimizer in the virtual environment as described in the documentation does not copy the file minimizer/grid/slurm-array-job.template. This leads to the provided example grid job (issue335_PDDL/grid_test.py) to yield an error.