This paper presents the first use of graph neural networks (GNNs) for higher-order proof search and demonstrates that GNNs can improve upon state-of-the-art results in this domain. Interactive, higher-order theorem provers allow for the formalization of most mathematical theories and have been shown to pose a significant challenge for deep learning. Higher-order logic is highly expressive and, even though it is well-structured with a clearly defined grammar and semantics, there still remains no well-established method to convert formulas into graph-based representations. In this paper, we consider several graphical representations of higher-order logic and evaluate them against the HOList benchmark for higher-order theorem proving.

We present an environment, benchmark, and deep learning driven automated theorem prover for higher-order logic. Higher-order interactive theorem provers enable the formalization of arbitrary mathematical theories and thereby present an interesting challenge for deep learning. We provide an open-source framework based on the HOL Light theorem prover that can be used as a reinforcement learning environment. HOL Light comes with a broad coverage of basic mathematical theorems on calculus and the formal proof of the Kepler conjecture, from which we derive a challenging benchmark for automated reasoning approaches. We also present a deep reinforcement learning driven automated theorem prover, DeepHOL, that gives strong initial results on this benchmark.

Deep learning techniques lie at the heart of several significant AI advances in recent years including object recognition and detection, image captioning, machine translation, speech recognition and synthesis, and playing the game of Go. Automated first-order theorem provers can aid in the formalization and verification of mathematical theorems and play a crucial role in program analysis, theory reasoning, security, interpolation, and system verification. Here we suggest deep learning based guidance in the proof search of the theorem prover E. We train and compare several deep neural network models on the traces of existing ATP proofs of Mizar statements and use them to select processed clauses during proof search. We give experimental evidence that with a hybrid, two-phase approach, deep learning based guidance can significantly reduce the average number of proof search steps while increasing the number of theorems proved. Using a few proof guidance strategies that leverage deep neural networks, we have found first-order proofs of 7.36% of the first-order logic translations of the Mizar Mathematical Library theorems that did not previously have ATP generated proofs. This increases the ratio of statements in the corpus with ATP generated proofs from 56% to 59%.

We introduce differential refinement logic (dRL), a logic with first-class support for refinement relations on hybrid systems, and a proof calculus for verifying such relations. dRL simultaneously solves several seemingly different challenges common in theorem proving for hybrid systems: 1. When hybrid systems are complicated, it is useful to prove properties about simpler and related subsystems before tackling the system as a whole. 2. Some models of hybrid systems can be implementation-specific. Verification can be aided by abstracting the system down to the core components necessary for safety, but only if the relations between the abstraction and the original system can be guaranteed. 3. One approach to taming the complexities of hybrid systems is to start with a simplified version of the system and iteratively expand it. However, this approach can be costly, since every iteration has to be proved safe from scratch, unless refinement relations can be leveraged in the proof. 4. When proofs become large, it is difficult to maintain a modular or comprehensible proof structure. By using a refinement relation to arrange proofs hierarchically according to the structure of natural subsystems, we can increase the readability and modularity of the resulting proof. dRL extends an existing specification and verification language for hybrid systems (differential dynamic logic, dL) by adding a refinement relation to directly compare hybrid systems. This paper gives a syntax, semantics, and proof calculus for dRL. We demonstrate its usefulness with examples where using refinement results in easier and better-structured proofs.

Abstract This paper is a tutorial on how to model hybrid systems as hybrid programs in differential dynamic logic and how to prove complex properties about these complex hybrid systems in KeYmaera, an automatic and interactive formal verification tool for hybrid systems. Hybrid systems can model highly nontrivial controllers of physical plants, whose behaviors are often safety critical such as trains, cars, airplanes, or medical devices. Formal methods can help design systems that work correctly. This paper illustrates how KeYmaera can be used to systematically model, validate, and verify hybrid systems. We develop tutorial examples that illustrate challenges arising in many real-world systems. In the context of this tutorial, we identify the impact that modeling decisions have on the suitability of the model for verification purposes. We show how the interactive features of KeYmaera can help users understand their system designs better and prove complex properties for which the automatic prover of KeYmaera still takes an impractical amount of time. We hope this paper is a helpful resource for designers of embedded and cyber–physical systems and that it illustrates how to master common practical challenges in hybrid systems verification.

As airspace becomes ever more crowded, air traffic management must reduce both space and time between aircraft to increase throughput, making on-board collision avoidance systems ever more important. These safety-critical systems must be extremely reliable, and as such, many resources are invested into ensuring that the protocols they implement are accurate. Still, it is challenging to guarantee that such a controller works properly under every circumstance. In tough scenarios where a large number of aircraft must execute a collision avoidance maneuver, a human pilot under stress is not necessarily able to understand the complexity of the distributed system and may not take the right course, especially if actions must be taken quickly. We consider a class of distributed collision avoidance controllers designed to work even in environments with arbitrarily many aircraft or UAVs. We prove that the controllers never allow the aircraft to get too close to one another, even when new planes approach an in-progress avoidance maneuver that the new plane may not be aware of. Because these safety guarantees always hold, the aircraft are protected against unexpected emergent behavior which simulation and testing may miss. This is an important step in formally verified, flyable, and distributed air traffic control.