Skip to content

Latest commit

 

History

History
62 lines (52 loc) · 2.6 KB

SAMPLING.md

File metadata and controls

62 lines (52 loc) · 2.6 KB
Raw

Sampling

Sampling Techniques

  • Code: src/search/sampling_techniques

  • Base class: SamplingTechnique

  • Purpose: Generate from a given task a new task (e.g. by changing the initial state)

  • Examples:

    • Modify the initial state by progression random walk: iforward_none(NB_OF_TASKS_TO_GENERATE, distribution=uniform_int_dist(MIN_WALK_LENGTH, MAX_WALK_LENGTH))

    • Generate a new initial state by regression random walk from the goal and afterwards completing the partial assignment to a state: backward_none(NB_OF_TASKS_TO_GENERATE, distribution=uniform_int_dist(MIN_WALK_LENGTH, MAX_WALK_LENGTH))

    • Please consult the code to see the full list of implemented techniques and all their parameters

Sampling Engines

  • Code: src/search/sampling_engines
  • Purpose: Use the sampling techniques to produce new tasks and extract training data from those tasks.

Examples:

  • Generate two new states via regression walks from the goal with walk lengths between 5 and 10. The states are solved using A*(LMcut). By default the samples are written to sas_plan. Ignore the message that no solution was found and check instead the line Generated Entries: [NUMBER]
./fast-downward.py --build debug ../benchmarks/gripper/prob01.pddl --search 
"sampling_search_simple(astar(lmcut(transform=sampling_transform()),
transform=sampling_transform()), techniques=[gbackward_none(2, 
distribution=uniform_int_dist(5, 10))])"

For all SamplingEngines and all parameters, take a look at the code. Below are the most important ones.

SamplingEngine

  • src/search/sampling_engines/sampling_engine.{h, cc}
  • manages the sampling techniques
    • queries them for new tasks, shuffles their order if desired, and stops once all SamplingTechniques are exhausted.
  • passes the new task to the abstract sample(task) method of its subclass and expects a list of strings (aka samples) back.
  • manages the samples (when and where to store them)

SamplingSearchBase : SamplingEngine

  • src/search/sampling_engines/sampling_search_base.{h, cc}
  • receives as additional parameter a search engine configuration (predefinitions are possible). For every new task, it creates the search engine object anew and executes the search. An abstract method extracts the samples from the search engine object.

SimpleSamplingSearch : SamplingSearchBase

  • src/search/sampling_engines/sampling_search_simple.{h, cc}
  • This engine extracts from the plan of an successful search
    • the states along the plan
    • the remaining path cost for the state
    • the operators used