On Compiling Away PDDL3 Soft Trajectory Constraints without Using Automata

Abstract

We address the problem of propositional planning extended with the class of soft temporally extended goals supported in PDDL3, also called qualitative preferences since IPC-5. Such preferences are useful to characterise plan quality by allowing the user to express certain soft constraints on the state trajectory of the desired solution plans. We propose and evaluate a compilation approach that extends previous work on compiling soft reachability goals and always goals to the full set of PDDL3 qualitative preferences. This approach directly compiles qualitative preferences into propositional planning without using automata to represent the trajectory constraints. Moreover, since no numeric fluent is used, it allows many existing STRIPS planners to immediately address planning with preferences without changing their algorithms or code. An experimental analysis presented in the paper evaluates the performance of state-of-the-art propositional planners using our compilation of qualitative preferences. The results indicate that the proposed approach is highly competitive with respect to current planners that natively support the considered class of preferences, as well as with a recent automata-based compilation approach.

Translations from Discretised PDDL+ to Numeric Planning

“…A wellknown class of reformulation approaches aims at translating a model from the original input language, to a different one. The idea is usually to remove the use of some poorly supported features of the language (see for instance (Ceriani and Gerevini 2015;Percassi and Gerevini 2019)) or to re-represent the problem in a less expressive language. The latter strategy has the advantage of increasing the number of planning engines that are able to reason upon the planning problem, at the cost of making the model extremely hard to read for human experts.…”

On Planning with Qualitative State-Trajectory Constraints in PDDL3 by Compiling them Away

“…For each instance with preferences, we ran some planners supporting preferences, i.e. LAMA and Mercury (Domshlak, Hoffmann, and Katz 2015) with the compiler by Percassi and Gerevini (2019), and LPRPG-P by Coles and Coles (2011), and we collected all plans generated within 30 CPU minutes. Out of this set, we took up to 5 plans (those with the larger number of satisfied preferences) and, for each of them, we generated a new problem instance having all satisfied preferences converted in qualitative state-trajectory constraints.…”

“…We propose a novel compilation-based schema that is specifically designed for qualitative PDDL3.0 trajectory constraints. Our compilation substantially revisits the automataless approach proposed by Percassi and Gerevini (2019) for soft trajectory constraints through the lens of hard trajectory constraints. We do so by revising the schema in two main important aspects: first, we characterize the state-trajectory constraints into two classes.…”

Reformulation Techniques for Automated Planning: A Systematic Review

“…The main advantage of this approach is increasing the number of planning engines that are able to reason upon the planning problem, and leverage existing robust technologies devised for solving more restricted cases. Examples of this class of reformulation approaches include [Percassi et al, 2021], that translates PDDL+ problems into PDDL2.1 ones, [Grastien andScala, 2020, Taig andBrafman, 2013] that provides approaches for translating of conformant planning problems into classical problems, the rerepresentation of uncertainty in conformant planning problems Palacios and Geffner [2009], the translation of complex temporal aspects in PDDL2.1 [Cooper et al, 2010], and the removal from PDDL3 of soft trajectory constraints [Percassi and Gerevini, 2019].…”