AoC number
Primary domain
Secondary domain
Description
There are two kinds of artificial intelligence: “weak” and “strong” AI. Weak AI is designed to perform a specific task at a human level, but can only perform its given task. Strong AI, meanwhile, has all of the intellectual capacity and processing power of a human brain, and can learn beyond its programming through observation and recalculation of its algorithms. Thanks to advances in neural networks, as seen in game-playing systems such as DeepStack and Libratus, the capacities of artificial intelligence have increased. Aviation is currently experimenting with strong AI, including systems capable of independent machine learning.
Complex engineered products are more likely to meet performance requirements when NDA are used. Aircraft structural health management has always relied upon NDA, with systems investigating root causes and identifying solutions. Management of the Next Generation Air Transportation System (NextGen) will use NDA for trajectory-based operations (TBO) to account for aircraft position and weather uncertainty. Carbonell is working on an artificial intelligence system that can identify holes in aircraft security, cross-check references from multiple aircraft, and dig for data to solve the issue, all autonomously.
Future flight decks may contain, or be expected to interact with, software “intelligent agents.” The characteristics of these agents may differ significantly from most software tools in use today. The increasing complexity of technology drives the need for such NDA. Entities such as Baomar and DARPA are testing intelligent autopilot systems, with DARPA’s ALIAS project landing and flying a simulated Boeing 737 in isolation.
Potential hazard
- Certification challenges due to non-deterministic nature of AI outputs from integrated modular architectures
- Pilots not understanding intent and actions of AI avionics
- Failure to achieve robustness, as defined under DO-178B guidelines – the very specific proof that under all application failure conditions, a single failure in one partition will not affect other partitions.
Corroborating sources and comments
“Nondeterministic approaches,” AIAA Aerospace America, December 2011
http://www.cotsjournalonline.com/articles/view/101451
Collision avoidance in commercial aircraft Free Flight via neural networks and non-linear programming, Christodoulou MA, Kontogeorgou C
http://www.ncbi.nlm.nih.gov/pubmed/18991361
http://www.theuav.com/; Autonomy is commonly defined as the ability to make decisions without human intervention. To that end, the goal of autonomy is to teach machines to be “smart” and act more like humans. Past efforts in the field of artificial intelligence include expert systems, neural networks, machine learning, natural language processing, and vision. To some extent, the ultimate goal in the development of autonomy technology is to replace the human pilot.
http://www.sciencemag.org/news/2017/03/artificial-intelligence-goes-deep-beat-humans-poker (Not aviation, but a recent advance in machine learning. DeepStack only calculates a few moves ahead of time and recalculates its algorithms based off of new information, and has defeated human players. Libratus uses three A.I. in conjunction – two to teach each other by playing multiple games, and a third to delete information from itself so it can’t be predicted in the long-term. Both can play imperfect information games successfully, setting them apart from similar A.I.)
https://www.wired.com/2017/01/mystery-ai-just-crushed-best-human-players-poker/ (Alternate take on above, focusing on Libratus rather than DeepStack. Note in particular that the human players who fought Libratus believed that it was learning from them as it played, and closed holes in its own strategy accordingly.)
https://www.wired.com/2017/03/ai-wields-power-make-flying-safer-maybe-even-pleasant/ (Overview of advanced A.I. and their potential roles in flight. Note in particular the possibility of systems that can identify holes, cross-check references from multiple aircraft, and dig for data to solve the issue; Carbonell is working on this now. Baomar is currently testing a more intelligent autopilot system on the X-Plane Flight Simulator, one of the most detailed on its kind, with the goal of creating a system that can respond to unexpected events.)
http://airfactsjournal.com/2016/10/artificial-intelligence-boom-means-aviation/ (As of 2016, University College London is working on a similar project to Baomar’s replacing the weak A.I. of a typical autopilot with a stronger, machine-learning system. The article also mentions DARPA’s ALIAS project, providing, essentially, an A.I. copilot. Mentions human-factors possibilities for interaction with A.I., including voice commands [as demonstrated by the Telligence system].)
http://www.airbusgroup.com/int/en/news-media/corporate-magazine/Forum-89/Artificial-Intelligence.html (Airbus’s take on the A.I. boom. Note in particular the possibility of an A.I. similar to IBM’s Watson, which can investigate root causes for itself and patch solutions – i.e. Libratus. Also, a system called TensorFlow is learning how to process images and identify objects for itself – a key feature for an aviation system. Currently free to download.)
Corroborating sources and comments
“Nondeterministic approaches,” AIAA Aerospace America, December 2011
http://www.cotsjournalonline.com/articles/view/101451
Collision avoidance in commercial aircraft Free Flight via neural networks and non-linear programming, Christodoulou MA, Kontogeorgou C
http://www.ncbi.nlm.nih.gov/pubmed/18991361
http://www.theuav.com/; Autonomy is commonly defined as the ability to make decisions without human intervention. To that end, the goal of autonomy is to teach machines to be “smart” and act more like humans. Past efforts in the field of artificial intelligence include expert systems, neural networks, machine learning, natural language processing, and vision. To some extent, the ultimate goal in the development of autonomy technology is to replace the human pilot.
http://www.sciencemag.org/news/2017/03/artificial-intelligence-goes-deep-beat-humans-poker (Not aviation, but a recent advance in machine learning. DeepStack only calculates a few moves ahead of time and recalculates its algorithms based off of new information, and has defeated human players. Libratus uses three A.I. in conjunction – two to teach each other by playing multiple games, and a third to delete information from itself so it can’t be predicted in the long-term. Both can play imperfect information games successfully, setting them apart from similar A.I.)
https://www.wired.com/2017/01/mystery-ai-just-crushed-best-human-players-poker/ (Alternate take on above, focusing on Libratus rather than DeepStack. Note in particular that the human players who fought Libratus believed that it was learning from them as it played, and closed holes in its own strategy accordingly.)
https://www.wired.com/2017/03/ai-wields-power-make-flying-safer-maybe-even-pleasant/ (Overview of advanced A.I. and their potential roles in flight. Note in particular the possibility of systems that can identify holes, cross-check references from multiple aircraft, and dig for data to solve the issue; Carbonell is working on this now. Baomar is currently testing a more intelligent autopilot system on the X-Plane Flight Simulator, one of the most detailed on its kind, with the goal of creating a system that can respond to unexpected events.)
http://airfactsjournal.com/2016/10/artificial-intelligence-boom-means-aviation/ (As of 2016, University College London is working on a similar project to Baomar’s replacing the weak A.I. of a typical autopilot with a stronger, machine-learning system. The article also mentions DARPA’s ALIAS project, providing, essentially, an A.I. copilot. Mentions human-factors possibilities for interaction with A.I., including voice commands [as demonstrated by the Telligence system].)
http://www.airbusgroup.com/int/en/news-media/corporate-magazine/Forum-89/Artificial-Intelligence.html (Airbus’s take on the A.I. boom. Note in particular the possibility of an A.I. similar to IBM’s Watson, which can investigate root causes for itself and patch solutions – i.e. Libratus. Also, a system called TensorFlow is learning how to process images and identify objects for itself – a key feature for an aviation system. Currently free to download.)