Combinatorial Artificial Intelligence for Defence Applications PhD

'First Wave AI' is used to describe the rules/logic based AI used heavily in the 1990's and 2000's and still in wide use today.  This involves 'handcrafted' expert systems, which are good at reasoning about narrowly defined problems, but poor at handling uncertainty and have no ability to learn or abstract/generalise. In that sense, these systems serve as complex functional approximators trained over an input-output data set.  

‘Second Wave AI’ is the term used to describe the current glut of 'machine learning' style intelligence, where algorithms are used that allow a computer to process large data-sets and learn patterns and behaviours, thus allowing them to respond when the same patterns are seen in new data.  This include 'supervised learning’ approaches (such as Deep CNN’s) and ‘unsupervised learning’ approaches (such as reinforcement based learning and generative adversarial networks).  Some of the main problems with Second Wave AI are 'explainability' and trust - as the machines learn, they are based upon statistical outcomes on large data sets, rather than human intuitive information.  Another problem lies with the fragility of the systems, 'illogical' outcomes can sometimes be generated due to biases, gaps or pollution of the training sets.  They typically lack the ability to generalise and to reason beyond what it has been trained over.  

It is an emerging opinion that the next advances will be achieved through combinations of these alternate approaches.  These may be loosely coupled (novel applications of existing techniques) or tightly coupled, which involves new ways of defining and developing these intelligences to combine both approaches. As such, recent advances on techniques such as Meta Learning, One-shot/Few-shot Learning and Distributed/Decentralized Federated Learning not only provide approaches to combine intelligence but also ensure computational tractability of exponentially growing and unbounded variable and instance sets. In addition, novel approaches such as Physics Informed/Guided Learning allows the learning models to capture the underlying physics/patterns and to generate physically consistent regression (or classification) which is applicable not only to the limited physical envelope of the data, but to a wider extend and thus generalise. Such approaches provide a balance between infinite extent models and limited extend data based on trust over particular sets, and naturally create explainable AI structures which can further be analysed from a verification and validation perspective. 

This research topic aims to define novel approaches to developing and combining these intelligences, utilizing both 1st and 2nd wave AI approaches, in the context of Defence applications. Such applications are expected to include: 

• Robust and “functionally explainable” machine-aided decision support for Safety and Mission Critical objectives e.g. fault detection/tracing, evasive manoeuvring, target selection etc. 
• Detailed semantic understanding of operational environments for Machine Situational Awareness, particularly within contested, congested and degraded scenarios. 
• Fully autonomous robust intelligence data processing to significantly reduce the reliance upon human analysts and counter huge increases in data volumes. 
• Improved synthetic training utilising machine-based instructors, matched to individual training needs. 
• Improved “Virtual Assistants” for the next generation of platform-operator interfaces.