Dr Nigel Greenwood
The Machine Genes Group
Focus: Evolutionary Artificial Intelligence (AI)
Fellowships awarded: 2005 and 2007
Importance to Australia: Of the dominant types of artificial intelligence (AI) in use today, my focus is on evolutionary AI. It is theoretically adept at solving problems relating to complex dynamic systems. It can evolve models of dynamical systems—missiles, aircraft, engines and diseased organisms, to name a few — from small, noise-polluted fragments of information and its results are fully explainable. However, the traditional forms of evolutionary AI suffered from two severe limitations: 'stagnation', where evolution stops working, and 'interpretation', where the evolutionary processes produce ambiguous results that are difficult to interpret or use. As a result of many years of research, I had solved these two problems.
Research overview: My research focuses on the development and use of evolutionary AI algorithms to solve complex problems in air-defence against sophisticated missiles. Other applications of my work include reducing operational risks and costs in aerospace and maritime engines, as well as solving difficult medication-control problems in medicine.
My earlier (1994) doctoral thesis in mathematics had addressed a related problem: the emerging danger of various technologies converging, namely supersonic cruise missiles with networked-enabled sensors and AI algorithms able to run on computing resources onboard the missile. Such missiles would not have predictable trajectories. This, combined with the nonlinear aerodynamics of their supersonic flight, would invalidate the assumptions made by conventional air-defence missiles. Did alternative air-defence approaches exist that were robust to these new technologies?
The Fellowship was awarded to address the consequent question: could the proposed solution outlined in my thesis be combined with my subsequent AI work, to build a viable air-defence architecture able to respond to these new, sophisticated threats? The necessary computing resources weren't available in the first decade of this century, but I was able to use the Fellowship to sketch out what the solution would look like. More recent work on high-performance computers has confirmed the validity of this approach.
Two other applications our Group is addressing using the same proprietary AI platforms are, first, military aviation engines. Complex dynamic systems (e.g. aircraft engines) are difficult and expensive to operate and maintain. Using evolutionary AI, we treat the digital twin of the engine as a living organism. These digital twins are extremely complicated, sophisticated physics-based models of the engine. The AI literally evolves and updates these digital twins from the engines' internal sensors in a simulated environment, using small fragments of sensor data. The AI then analyses these models to identify and address deterioration in the actual individual engines due to wear-and-tear before they become an operational problem. With each engine costing $US 13-35 million, the ability to address issues before they impair availability will be significant in reducing operational risk and cost.
The second application of our AI is the development of an artificial pancreas to deliver the safe dose of insulin for highly unstable diabetes cases, where a distinct model must be evolved for each patient from tiny amounts of medical data, and the patient's entire history is one of therapeutic failure, so the pattern-recognition techniques of conventional AI are useless.
What inspired you to pursue this area of research?
As an applied mathematician, I was fortunate to join an international team conducting world-leading research into novel forms of artificial intelligence, and to be mentored by the leaders of this team: first by the University of Queensland's Dr. Janislaw Skowronski and then by UC Berkeley's Emeritus Professor George Leitmann, who were working in industrial robotics.
Our research showed that AI was heading in the wrong direction. We showed that to be useful to industrial robotics and other complex, dynamic systems, it needed to be able to understand itself and the systems it was required to repair and maintain.
Like an aircraft engine, an industrial robotic arm is a complex and expensive piece of equipment. To keep the robotic arm operating effectively, the AI needs to get the robot arm to understand its internal operations and to self-correct, based on this understanding. After all, it may not be possible to understand a quirk in robotic behaviour by dismantling the robot, particularly as the operation of all the parts is tightly interconnected.
How does your research address existing gaps and challenges in your specialist field?
Complex problems are not impossible problems. Evolutionary AI gives us the capability to solve them, without the expensive computing resources and masses of data that are required by neural networks. Much is made of neural networks in AI and their use across a range of industries. However, the sheer amount of data needed, the requirement to 'clean' these data prior to use and the cost of the compute power makes neural networks an expensive proposition. There are also the issues of 'hallucination' and 'drift' in which the neural networks see patterns that do not exist and make decisions based on an incorrect analysis of these patterns.
The architecture of evolutionary AI that I have developed overcomes the issue of stagnation, where the evolutionary process stops making progress. I achieved this by reformulating how evolution should be described mathematically. The other problem, of interpretation, was solved when I created a form of generative adversarial AI, where one AI challenges and tries to exploit weaknesses put forward by another AI when trying to interpret a model. Imagine multiple experts debating solutions to difficult problems and generating new or improved solutions simultaneously in real-time, and 'learning' to come up with the best solution by challenging each other's proposals. This form of generative adversarial AI was successfully demonstrated in 2012, two years before generative adversarial networks were invented, and in 2014 patent applications were lodged. These have now been awarded.
How will your research contribute to Australian sovereign capabilities and defence innovation?
In the current Cold War, a global arms race exists to develop new forms of AI. The conventional orthodoxy is that success in this arms race requires access to huge amounts of data and onerous computing, hence requires access to large-scale supercomputing. Were this to be true, it would effectively exclude smaller countries like Australia from directly engaging in AI-based Defence innovation, but instead require us merely to acquire already-developed technologies from the United States.
In September 2021 Nicholas Chaillan, the USAF/US Space Force Chief Software Officer, took this argument a step further. He resigned, publicly announcing that 'China has won the AI battle with the US' because of the issues of data availability, skilled labour and computing resources.
By demonstrating that alternative forms of AI exist, built on completely different paradigms that successfully work with tiny fragments of noise-polluted sensor data and have much lower computational demands, my research has shown that Australia can directly engage in globally novel AI-based military innovation and build sovereign capability as a sovereign enterprise. Chaillan was wrong.
What does the Spitfire Memorial Defence Fellowship (SMDF) and the funding awarded to you mean to you and your research?
The Spitfire Memorial Defence Fellowships meant I could explore novel technologies and solutions, despite the fact these were radically different from what DSTO imagined at the time to be appropriate.
The result was I could design and build an early prototype of my eventual platform, enabling a more agile approach to AI-enhanced command and control of multiple different assets in a highly distributed environment.
The genius of the Spitfire Memorial Defence Fellowship is that it gives researchers complete freedom in how they conduct their research. It is not bound in excessive reporting and governance requirements. This means researchers and innovators can take novel approaches to difficult and ambiguous problems, to come up with more inventive and practical solutions, without being constrained by current thinking or bureaucratic orthodoxy.
I was also fortunate to hold the SMDF at a time when some of the original Spitfire pilots were still alive. They actively encouraged me to 'keep my chin up' when struggling with the bureaucracy.
The SMDF has the capacity to make a real difference to Australian sovereign capabilities and defence innovation. It reflects the spirit of the pilots, Reginald Joseph Mitchell, C.B.E., F.R.Ae.S. and the Spitfire design. Mitchell solved the design challenges with the Spitfire by looking at it the way a mathematician does, seeking radical elegance, and not as an engineering task based on incremental improvement. He looked at what could be achieved and was not limited by the thinking of the time, nor by the engineering challenges in building and constantly developing the Spitfire.