Towards Verification of Neural Networks for Small Unmanned Aircraft Collision Avoidance

Irfan, Ahmed; Julian, Kyle D.; Wu, Haoze; Barrett, Clark; Kochenderfer, Mykel J.; Meng, Baoluo; Lopez, James G.

doi:10.1109/dasc50938.2020.9256616

Cited by 9 publications

(3 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Julian and Kochenderfer [13] train a deep neural network to approximate the ACAS X lookup table to reduce the storage needs and runtime of the system, and Irfan et al [9] and Julian and Kochenderfer [12] explore applying formal methods to verify such neural networks. One drawback to this approach is that SMT does not support continuous dynamics, and all queries to the SMT solver must be in the form of discrete, linear regions.…”

Section: Related Workmentioning

confidence: 99%

Formally Verified Next-generation Airborne Collision Avoidance Games in ACAS X

Cleaveland

Mitsch

Platzer

2022

ACM Trans. Embed. Comput. Syst.

View full text Add to dashboard Cite

The design of aircraft collision avoidance algorithms is a subtle but important challenge that merits the need for provable safety guarantees. Obtaining such guarantees is nontrivial given the unpredictability of the interplay of the intruder aircraft decisions, the ownship pilot reactions, and the subtlety of the continuous motion dynamics of aircraft. Existing collision avoidance systems, such as TCAS and the Next-Generation Airborne Collision Avoidance System ACAS X, have been analyzed assuming severe restrictions on the intruder’s flight maneuvers, limiting their safety guarantees in real-world scenarios where the intruder may change its course. This work takes a conceptually significant and practically relevant departure from existing ACAS X models by generalizing them to hybrid games with first-class representations of the ownship and intruder decisions coming from two independent players, enabling significantly advanced predictive power. By proving the existence of winning strategies for the resulting Adversarial ACAS X in differential game logic, collision-freedom is established for the rich encounters of ownship and intruder aircraft with independent decisions along differential equations for flight paths with evolving vertical/horizontal velocities. We present three classes of models of increasing complexity: single-advisory infinite-time models, bounded time models, and infinite time, multi-advisory models. Within each class of models, we identify symbolic conditions and prove that there then always is a possible ownship maneuver that will prevent a collision between the two aircraft.

show abstract

Section: Related Workmentioning

confidence: 99%

Formally Verified Next-generation Airborne Collision Avoidance Games in ACAS X

Cleaveland

Mitsch

Platzer

2022

ACM Trans. Embed. Comput. Syst.

View full text Add to dashboard Cite

show abstract

“…Such methods aim to mathematically prove that a given system can(not) respond in particular ways to particular inputs (Katz et al, 2017b;a;Kuper et al, 2018;Katz et al, 2019). For instance, formal verification was used to study the safety of neural networks used for unmanned aircraft collision avoidance (Irfan et al, 2020). However, such methods remain largely untested for advanced AI models (Dalrymple et al, 2024).…”

Section: Verification Of Model Propertiesmentioning

confidence: 99%

AI Governance Multi-stakeholder Convening

Greene

2022

The Oxford Handbook of AI Governance

View full text Add to dashboard Cite

This chapter offers reflections and advice on AI ethics and governance from a year spent leading Partnership on AI’s multi-stakeholder Affective Computing and Ethics project, involving more than 200 engineers, scientists, lawyers, privacy and civil rights advocates, bioethicists, managers, executives, journalists, and government officials, mostly in the United States, United Kingdom, and European Union, in discussions about AI related to emotion and affect and its potential impacts on civil and human rights, with a goal of developing industry best practices and better technology policy. The author’s reflections from that year draw lessons on convening, multi-disciplinary collaboration, and affective computing and AI ethics and governance. The chapter offers a blueprint for creating the shared knowledge foundation non-technical participants need to apply their expertise. It presents question exploration as a tool for evaluating ethics risk and using “What is notice good for?” shows how a well-chosen question can serve as a catalyst for group or individual exploration of the issues, leading to insights and answers. It includes a curated list of 42 questions, catalysts for AI and ethics discussion for industry, university, news media, and policy teams; multi-disciplinary, multi-stakeholder AI ethics and governance convenings; and for individual writing and thinking, to take scholars and practitioners a step beyond or to the side of what they have been thinking about—about AI and ethics.

show abstract

“…Julian and Kochenderfer [11] train a deep neural network to approximate the ACAS X lookup table to reduce the storage needs and required runtime of the system, and Irfan, et al [12] and Julian and Kochenderfer [13] explore applying formal methods to verify such neural networks. One drawback to this approach is that SMT does not support representations of continuous dynamics, and all queries to the SMT solver must be in the form of discrete, linear regions.…”

Section: Related Workmentioning

confidence: 99%

Formally Verified Next-Generation Airborne Collision Avoidance Games in ACAS X

Cleaveland¹,

Mitsch²,

Platzer³

2021

Preprint

View full text Add to dashboard Cite

The design of aircraft collision avoidance algorithms is a subtle but important challenge that merits the need for provable safety guarantees. Obtaining such guarantees is nontrivial given the unpredictability of the interplay of the intruder aircraft decisions, the ownship pilot reactions, and the subtlety of the continuous motion dynamics of aircraft. Existing collision avoidance systems, such as TCAS and the Next-Generation Airborne Collision Avoidance System ACAS X, have been analyzed assuming severe restrictions on the intruder's flight maneuvers, limiting their safety guarantees in real-world scenarios where the intruder may also be changing its course.This work takes a conceptually significant and practically relevant departure from existing ACAS X models by generalizing them to hybrid games with first-class representations of the ownship and intruder decisions coming from two independent players. By proving the existence of winning strategies for the resulting Adversarial ACAS X in differential game logic, collision-freedom is established for the rich encounters of ownship and intruder aircraft with independent decisions along differential equations for flight paths with evolving vertical/horizontal velocities. We present three classes of models of increasing complexity: singleadvisory infinite-time models, bounded time models, and infinite time, multi-advisory models. Within each class of models, we identify symbolic conditions and prove that there then always is a possible ownship maneuver that will prevent a collision between the two aircraft. CCS Concepts: • Theory of computation → Timed and hybrid models; Modal and temporal logics; Programming logic; • Computer systems organization → Embedded systems.

show abstract

Towards Verification of Neural Networks for Small Unmanned Aircraft Collision Avoidance

Cited by 9 publications

References 9 publications

Formally Verified Next-generation Airborne Collision Avoidance Games in ACAS X

Formally Verified Next-generation Airborne Collision Avoidance Games in ACAS X

AI Governance Multi-stakeholder Convening

Formally Verified Next-Generation Airborne Collision Avoidance Games in ACAS X

Contact Info

Product

Resources

About