Learning Features and Abstract Actions for Computing Generalized Plans

Abstract

Generalized planning is concerned with the computation of plans that solve not one but multiple instances of a planning domain. Recently, it has been shown that generalized plans can be expressed as mappings of feature values into actions, and that they can often be computed with fully observable non-deterministic (FOND) planners. The actions in such plans, however, are not the actions in the instances themselves, which are not necessarily common to other instances, but abstract actions that are defined on a set of common features. The formulation assumes that the features and the abstract actions are given. In this work, we address this limitation by showing how to learn them automatically. The resulting account of generalized planning combines learning and planning in a novel way: a learner, based on a Max SAT formulation, yields the features and abstract actions from sampled state transitions, and a FOND planner uses this information, suitably transformed, to produce the general plans. Correctness guarantees are given and experimental results on several domains are reported.

A review of generalized planning

“…In this case the sets of possible initial and goal states are represented exclusively using literals and the three basic logical connectives ( and, to indicate a conjunction of literals or, to indicate a disjunction of literals and not, to indicate negation). Examples of sets of planning instances represented with propositional logic are conformant, contingent or POMDPs planning tasks that define the different possible initial states of the task as a disjunction on the problem literals (goals are shared for all the possible initial states in the planning task) (Bonet et al ., 2010). First-order logic .…”

“…conditional transitions) to keep the solution space tractable. This is what happens in FSCs (Bonet et al ., 2010), generalized policies (Martn and Geffner, 2004), or with the conditional gotos used in planning programs (Segovia-Aguas et al ., 2016a).…”

“…Some forms of generalized plans include explicit termination actions (or states) so goal testing is only done when such actions are applied (or such states reached) (Srivastava et al ., 2011b; Jiménez and Jonsson, 2015). If the generalized plan lacks these termination actions (or states) goal achievement has to be tested after executing every plan step (Bonet et al ., 2010). The plan is incomplete .…”

“…Examples of this approach are compilations of generalized planning into other forms of problem solving such as classical planning (Jiménez and Jonsson, 2015), conformant planning (Bonet et al ., 2010), CSP (Pralet et al ., 2010) or a Prolog program (Hu and Giacomo, 2013). These compilations implement a search space as described above, and benefit from off-the-shelf solvers (with efficient search algorithms and heuristics) to complete the search for a generalized plan.…”

“…Each form with its own expressiveness capacity and own computation and validation complexity. Generalized plans range from programs (Winner and Veloso, 2003; Segovia-Aguas et al ., 2016a) and generalized polices (Martn and Geffner, 2004) to Finite State Controllers (FSCs) (Bonet et al ., 2010; Segovia-Aguas et al ., 2016b), AND/OR graphs, formal grammars (Ramirez and Geffner, 2016) or hierarchical task networks (HTNs) (Nau et al ., 2003). We can classify generalized plans according to their specification of the action to apply next : Fully specified solutions, that unambiguously specify the action to apply next , for solving every instance in a given generalized planning task.…”