Differences between revisions 12 and 13
Revision 12 as of 2022-12-09 09:36:36
Size: 7663
Comment: add new long-term plans
Revision 13 as of 2023-07-28 10:18:08
Size: 7666
Editor: MalteHelmert
Comment:
Deletions are marked like this. Additions are marked like this.
Line 60: Line 60:
To fully support task transformations, we have to be aware of all interactions between different levels. The main idea is to have a hierarchy of task transformations with a {{{RootTask}}} at the root and transformations as edges/children in this hierarchy. As a first step, the search engines will always use the root task but later we want to support transforming the task before giving it to the search. Heuristics can transform the task they get from the search and some of them explicitly wrap the task in some transformation (e.g., CEGAR uses {{{DomainAbstractedTask}}} for the landmark decomposition). We thus have two interfaces that cross levels in the hierarchy: To fully support task transformations, we have to be aware of all interactions between different levels. The main idea is to have a hierarchy of task transformations with a {{{RootTask}}} at the root and transformations as edges/children in this hierarchy. As a first step, the search algorithms will always use the root task but later we want to support transforming the task before giving it to the search. Heuristics can transform the task they get from the search and some of them explicitly wrap the task in some transformation (e.g., CEGAR uses {{{DomainAbstractedTask}}} for the landmark decomposition). We thus have two interfaces that cross levels in the hierarchy:

Back to ForDevelopers.

Long-term plans (December 2022)

Planner as a library

  • easier application for many users
  • part of AIplan4EU
  • required changes:
    • single process planner, no (intermediate) file writing, possibly via Python bindings of C++ modules
    • solve component interaction problem
    • get rid of the last global data structures (used in search)
    • re-write translator in C++

Integrate lifted techniques

  • promised as part of recently granted funding
  • requires making state representation and search techniques more general

Large planner extensions

  • need to discuss and decide which large features to implement or to abandon
  • examples:
    • cost partitioning
    • symmetries
    • certificates

Long-term plans (older - TODO: remove)

  • state representation (issue401)

  • life-time management of shared objects (e.g. CG) (related: issue564)

  • rethink which shared objects can go together (e.g. when is it OK for a heuristic to use a different task than a search algorithm using it? When is it OK for a pattern selector to work with a different task? a landmark factory? an operator-counting constraint generator? etc.)

Needs tender love and care

  • landmarks
  • option parser
  • search

Items with action plans

Component creation and interaction

Terminology and action plan:

  • component: e.g. FFHeuristic, AlternationOpenList, LandmarkFactorySasp

  • component builder: unpacks and stores options and creates components
  • plugin: creates a component builder, interacts with option parser

We implemented this action plan in a prototype (see http://issues.fast-downward.org/issue559). It currently doesn't include the following ideas, which we had during an offline discussion:

  • Use classes for plugins instead of _parse functions.

  • Plugin classes allow adding commandline options, call OptionParser::parse(), create component builders (and pass the Options object to their constructor).

  • if needed a common plugin base class could handle things like dry_run()

Task Transformations

To fully support task transformations, we have to be aware of all interactions between different levels. The main idea is to have a hierarchy of task transformations with a RootTask at the root and transformations as edges/children in this hierarchy. As a first step, the search algorithms will always use the root task but later we want to support transforming the task before giving it to the search. Heuristics can transform the task they get from the search and some of them explicitly wrap the task in some transformation (e.g., CEGAR uses DomainAbstractedTask for the landmark decomposition). We thus have two interfaces that cross levels in the hierarchy:

Search-Heuristic interface

The search has to interact with the heuristic in the following way:

  • States from the search space of the search have to be translated down the hierarchy into states in the search space of the heuristic task.
    • Implementation: The transformation is done inside the heuristic in Heuristic::convert_global_state by calling task_proxy.convert_ancestor_state(parent_state_data) where task_proxy is a proxy for the task used in the heuristic and parent_state_data is the data of the search state. The actual transformation is done in AbstractTask::convert_state_values which has to be implemented when designing a new task transformation. The default implementation in DelegatingTask passes the call to the parent to get a state of the parent task, then calls convert_state_values_from_parent to convert it to a local state. If any of the methods is not implemented or the input state is not from an ancestor state on the hierarchy, the call fails at the latest when reaching the root of the hierarchy.

    • Assumptions: we assume that the task used in the heuristic is a descendant in the hierarchy from the task used in the search. If this is not the case, converting states will fail with an error.

  • Preferred operators identified by a heuristic have to be translated up the hierarchy to operators of the task used in the search.
    • Implementation: Heuristic::set_preferred translates all operator IDs from local IDs to IDs of the root task before inserting them into the set of preferred operators that is handed to the search. This uses OperatorProxy::get_global_operator_id which is handled in AbstractTask::get_global_operator_id. Currently, none of our transformations touches the operators, so the call is forwarded to the root task where the function is the identity function, and then back down the hierarchy without change. Newly implemented task transformations that change the operator have to override the function with an actual translation.

    • Issues: we currently do not support translating operators back up only part of the way in the hierarchy which would be needed if the search can work on a transformed task. We also currently do not support returning a set of operators that correspond to a single operator which might be useful if the task transformation combines several operators into one.

    • Assumptions: we assume that the task used in the heuristic can identify a single operator of the root task that corresponds to a given operator. If this is not the case, the task transformation should abort in the method.

  • Transitions in the search space are passed on to path-dependent evaluators using the methods notify_initial_state and notify_state_transition. They are transitions (states and operators) of the search task and may have to be translated down the hierarchy. The only evaluator that currently uses it is the landmark count heuristic. Here, we have to either translate landmarks up the hierarchy (computing landmarks on a transformed task and translating the resulting set of landmarks into a set of landmarks for the search task) or we have to translate transitions of the search task down the hierarchy (computing landmarks on the transformed task, then checking if a new landmark is reached also in the transformed task).

    • Issues: none of these alternatives is implemented and it is not clear which of them is better. Depending on whether the transformed task is larger (e.g., Pi^m) or smaller (e.g., an abstraction) one or the other is what we want. At the time of this writing, task transformation are not correctly supported in the landmark code: the landmarks are computed on a transformed task but the transitions coming from the search are not translated, i.e., we treat the landmarks of the transformed task as landmarks of the search task. The transformation given to the landmark heuristic is wrapped in a CostAdaptedTask which in turn ignores the transformation. There are two independent issues here: CostAdaptedTask should not ignore its parameter (easy to fix now that is_unit_cost is no longer limited to the root task); and the landmark heuristic should either translate transitions or landmarks as described above.

User-Search interface

Currently, the search always uses the root task. If we want to change this, we have to consider the interface between the search and the user code. I currently only see one interaction there:

  • The plan returned by the search is a plan for the transformed task. We could offer methods to transform it back on demand.

FastDownward: ForDevelopers/LongTermPlans (last edited 2023-07-28 10:18:08 by MalteHelmert)