A Formally Verified Validator for Classical Planning Problems and Solutions

Abstract

In this paper we present a formally verified validator for planning problems and their solutions. We formalise the semantics of a fragment of PDDL (∨, ¬, →, = in the preconditions, typing and constants) in the Higher-Order Logic theorem prover Isabelle/HOL. We then construct an efficient plan validator and mechanically prove it correct w.r.t. our semantics. We argue that our approach provides a superior compromise in constructing validators where one can have the best of two worlds: (i) clear and concise semantics w.r.t. which the validator is built thus helping to avoid bugs (unlike existing validators, which we show have bugs) and (ii) an optimised implementation whose performance is competitive with mainstream unverified validators.

Formally Verified Solution Methods for Markov Decision Processes

“…Despite the impressive advances in the capabilities of different types of AI systems, it is becoming clear that one major hurdle to their wide adoption is the lack of trustworthiness of these systems. This has prompted researchers to study techniques to boost the trustworthiness of AI systems in different areas, like machine learning (Selsam, Liang, and Dill 2017;Katz et al 2017), planning (Abdulaziz, Gretton, and Norrish 2019;Abdulaziz and Lammich 2018), and model-checking (Esparza et al 2013). However, one of the areas of AI where there is still a lot to be done regarding trustworthiness is software for solving Markov decision processes (MDPs).…”

The Power of Reformulation: From Validation to Planning in PDDL+

“…A different line of research looked into the reformulation of the validation task into theorem proving (Abdulaziz and Lammich 2018). This has the advantage of reducing the possibility of bugs but is currently limited to classical planning formalism.…”

Formal Semantics and Formally Verified Validation for Temporal Planning

“…Although, performance-wise, planning algorithms and systems are very scalable and efficient, as shown by different planning competitions (Long et al 2000;Coles et al 2012;Vallati et al 2015), there is still to be desired when it comes to their trustworthiness, which is crucial to their wide adoption. Consequently, there have been substantial efforts to improve the trustworthiness of planning systems (Howey, Long, and Fox 2004;Fox, Howey, and Long 2005;Eriksson, Röger, and Helmert 2017;Abdulaziz, Norrish, and Gretton 2018;Abdulaziz and Lammich 2018;Cimatti, Micheli, and Roveri 2017;Abdulaziz, Gretton, and Norrish 2019). A basic task when it comes to the trustworthiness of planning systems is that of plan validation.…”