Martin Suda

Jan Jakubův, Karel Chvalovský, Zarathustra Goertzel et al.

As a present to Mizar on its 50th anniversary, we develop an AI/TP system that automatically proves about 60\% of the Mizar theorems in the hammer setting. We also automatically prove 75\% of the Mizar theorems when the automated provers are helped by using only the premises used in the human-written Mizar proofs. We describe the methods and large-scale experiments leading to these results. This includes in particular the E and Vampire provers, their ENIGMA and Deepire learning modifications, a number of learning-based premise selection methods, and the incremental loop that interleaves growing a corpus of millions of ATP proofs with training increasingly strong AI/TP systems on them. We also present a selection of Mizar problems that were proved automatically.

Lasse Blaauwbroek, David Cerna, Thibault Gauthier et al.

Automated theorem provers and formal proof assistants are general reasoning systems that are in theory capable of proving arbitrarily hard theorems, thus solving arbitrary problems reducible to mathematics and logical reasoning. In practice, such systems however face large combinatorial explosion, and therefore include many heuristics and choice points that considerably influence their performance. This is an opportunity for trained machine learning predictors, which can guide the work of such reasoning systems. Conversely, deductive search supported by the notion of logically valid proof allows one to train machine learning systems on large reasoning corpora. Such bodies of proof are usually correct by construction and when combined with more and more precise trained guidance they can be boostrapped into very large corpora, with increasingly long reasoning chains and possibly novel proof ideas. In this paper we provide an overview of several automated reasoning and theorem proving domains and the learning and AI methods that have been so far developed for them. These include premise selection, proof guidance in several settings, AI systems and feedback loops iterating between reasoning and learning, and symbolic classification problems.

Tomáš Balyo, Martin Suda, Lukáš Chrpa et al.

Existing planning action domain model acquisition approaches consider different types of state traces from which they learn. The differences in state traces refer to the level of observability of state changes (from full to none) and whether the observations have some noise (the state changes might be inaccurately logged). However, to the best of our knowledge, all the existing approaches consider state traces in which each state change corresponds to an action specified by its name and all its parameters (all objects that are relevant to the action). Furthermore, the names and types of all the parameters of the actions to be learned are given. These assumptions are too strong. In this paper, we propose a method that learns action schema from state traces with fully observable state changes but without the parameters of actions responsible for the state changes (only action names are part of the state traces). Although we can easily deduce the number (and names) of the actions that will be in the learned domain model, we still need to deduce the number and types of the parameters of each action alongside its precondition and effects. We show that this task is at least as hard as graph isomorphism. However, our experimental evaluation on a large collection of IPC benchmarks shows that our approach is still practical as the number of required parameters is usually small. Compared to the state-of-the-art learning tools SAM and Extended SAM our new algorithm is able to provide better results in multiple domains in terms of learning action models more similar to reference models, even though it uses less information and has fewer restrictions on the input traces.

Martin Suda

Clause selection is arguably the most important choice point in saturation-based theorem proving. Framing it as a reinforcement learning (RL) task is a way to challenge the human-designed heuristics of state-of-the-art provers and to instead automatically evolve -- just from prover experiences -- their potentially optimal replacement. In this work, we present a neural network architecture for scoring clauses for clause selection that is powerful yet efficient to evaluate. Following RL principles to make design decisions, we integrate the network into the Vampire theorem prover and train it from successful proof attempts. An experiment on the diverse TPTP benchmark finds the neurally guided prover improve over a baseline strategy, from which it initially learns -- in terms of the number of in-training-unseen problems solved under a practically relevant, short CPU instruction limit -- by 20%.

Filip Bártek, Karel Chvalovský, Martin Suda

To achieve the best performance, automatic theorem provers often rely on schedules of diverse proving strategies to be tried out (either sequentially or in parallel) on a given problem. In this paper, we report on a large-scale experiment with discovering strategies for the Vampire prover, targeting the FOF fragment of the TPTP library and constructing a schedule for it, based on the ideas of Andrei Voronkov's system Spider. We examine the process from various angles, discuss the difficulty (or ease) of obtaining a strong Vampire schedule for the CASC competition, and establish how well a schedule can be expected to generalize to unseen problems and what factors influence this property.

Martin Suda

We re-examine the topic of machine-learned clause selection guidance in saturation-based theorem provers. The central idea, recently popularized by the ENIGMA system, is to learn a classifier for recognizing clauses that appeared in previously discovered proofs. In subsequent runs, clauses classified positively are prioritized for selection. We propose several improvements to this approach and experimentally confirm their viability. For the demonstration, we use a recursive neural network to classify clauses based on their derivation history and the presence or absence of automatically supplied theory axioms therein. The automatic theorem prover Vampire guided by the network achieves a 41% improvement on a relevant subset of SMT-LIB in a real time evaluation.

Martin Suda

Vampire has been for a long time the strongest first-order automatic theorem prover, widely used for hammer-style proof automation in ITPs such as Mizar, Isabelle, HOL, and Coq. In this work, we considerably improve the performance of Vampire in hammering over the full Mizar library by enhancing its saturation procedure with efficient neural guidance. In particular, we employ a recently proposed recursive neural network classifying the generated clauses based only on their derivation history. Compared to previous neural methods based on considering the logical content of the clauses, our architecture makes evaluating a single clause much less time consuming. The resulting system shows good learning capability and improves on the state-of-the-art performance on the Mizar library, while proving many theorems that the related ENIGMA system could not prove in a similar hammering evaluation.

Jan Jakubův, Karel Chvalovský, Miroslav Olšák et al.

We describe an implementation of gradient boosting and neural guidance of saturation-style automated theorem provers that does not depend on consistent symbol names across problems. For the gradient-boosting guidance, we manually create abstracted features by considering arity-based encodings of formulas. For the neural guidance, we use symbol-independent graph neural networks (GNNs) and their embedding of the terms and clauses. The two methods are efficiently implemented in the E prover and its ENIGMA learning-guided framework. To provide competitive real-time performance of the GNNs, we have developed a new context-based approach to evaluation of generated clauses in E. Clauses are evaluated jointly in larger batches and with respect to a large number of already selected clauses (context) by the GNN that estimates their collectively most useful subset in several rounds of message passing. This means that approximative inference rounds done by the GNN are efficiently interleaved with precise symbolic inference rounds done inside E. The methods are evaluated on the MPTP large-theory benchmark and shown to achieve comparable real-time performance to state-of-the-art symbol-based methods. The methods also show high complementarity, solving a large number of hard Mizar problems.

Karel Chvalovský, Jan Jakubův, Martin Suda et al.

We describe an efficient implementation of clause guidance in saturation-based automated theorem provers extending the ENIGMA approach. Unlike in the first ENIGMA implementation where fast linear classifier is trained and used together with manually engineered features, we have started to experiment with more sophisticated state-of-the-art machine learning methods such as gradient boosted trees and recursive neural networks. In particular the latter approach poses challenges in terms of efficiency of clause evaluation, however, we show that deep integration of the neural evaluation with the ATP data-structures can largely amortize this cost and lead to competitive real-time results. Both methods are evaluated on a large dataset of theorem proving problems and compared with the previous approaches. The resulting methods improve on the manually designed clause guidance, providing the first practically convincing application of gradient-boosted and neural clause guidance in saturation-style automated theorem provers.

Giles Reger, Martin Suda, Andrei Voronkov

This paper attempts to address the question of how best to assure the correctness of saturation-based automated theorem provers using our experience developing the theorem prover Vampire. We describe the techniques we currently employ to ensure that Vampire is correct and use this to motivate future challenges that need to be addressed to make this process more straightforward and to achieve better correctness guarantees.

Giles Reger, Martin Suda, Andrei Voronkov et al.

Modern saturation-based Automated Theorem Provers typically implement the superposition calculus for reasoning about first-order logic with or without equality. Practical implementations of this calculus use a variety of literal selections and term orderings to tame the growth of the search space and help steer proof search. This paper introduces the notion of lookahead selection that estimates (looks ahead) the effect on the search space of selecting a literal. There is also a case made for the use of incomplete selection functions that attempt to restrict the search space instead of satisfying some completeness criteria. Experimental evaluation in the \Vampire\ theorem prover shows that both lookahead selection and incomplete selection significantly contribute to solving hard problems unsolvable by other methods.

Martin Suda

We describe a duality mapping between STRIPS planning tasks. By exchanging the initial and goal conditions, taking their respective complements, and swapping for every action its precondition and delete list, one obtains for every STRIPS task its dual version, which has a solution if and only if the original does. This is proved by showing that the described transformation essentially turns progression (forward search) into regression (backward search) and vice versa. The duality sheds new light on STRIPS planning by allowing a transfer of ideas from one search approach to the other. It can be used to construct new algorithms from old ones, or (equivalently) to obtain new benchmarks from existing ones. Experiments show that the dual versions of IPC benchmarks are in general quite difficult for modern planners. This may be seen as a new challenge. On the other hand, the cases where the dual versions are easier to solve demonstrate that the duality can also be made useful in practice.