ModPlan - Modern Action Planning
Domain Inference
Learning PDDL domain specifications from plan traces
without any domain expert knowledge is a
computationally challenging and practically infeasible task.
To infer operators within a PDDL domain description,
the inference mechanism needs supervision of the domain expert.
Knowledge Acquisition and Engineering
We newly implemented a supervised learning algorithm to interactively
infer the PDDL domain description.
We assume that a domain expert tries to infer a valid domain
description from a set of operators that form a valid plan. This
plan can be generated in a previous run of a planner. If we start from
scratch, an initial sequence of operators has to be provided manually.
The additional inputs of the algorithm are the
prefixes of domain and problem file,
namely the declaration of objects,
and the set of predicates.
If not already present, object type information may
interactively be attached.
Given the set of operators in a valid plan, the designer is confronted
with choice boxes on how the set of preconditions and effects of an
operator to be inferred are composed. The supervised PDDL
learning mechanism selects the operator to infer next and
steadily reduces the set of options until a domain
model has been established.
For long plans, the inference task is
almost fully automatic.
The learning algorithm underneath provides an slightly improved implementation
of the Opmaker algorithm. One of the distinctive
features is the option to attach durations
to actions and to allow incremental
learning, as the output of a planner can be used as input for
another inference step.