AI Insider: Integrating expert knowledge

Knowledge transfer from humans to AI systems, making them smarter, safer, more reliable and more secure.

Professor Michael Thielscher in the School of Computer Science and Engineering has been developing AI for more than 25 years.

“I know that everyone does AI now, but I've been doing it for a little longer,” he said.

“Essentially my work is on the more fundamental side – what we call ‘knowledge engineering’ or ‘reasoning and planning’. By integrating expert knowledge into our AI systems, we can make them smarter as well as safer, more reliable and more secure.”

Prof. Thielscher is currently collaborating on two “very interesting” cross-disciplinary projects with colleagues in UNSW Arts, Design & Architecture (ADA).

Urban Copilot with the City Futures Research Centre

The School of Built Environment’s City Futures Research Centre (CFRC) has developed a digital assistant for planners, developers and government staff called Urban Copilot, built on the success of a previous project MapAI. Urban Copilot simplifies the regulatory maze of urban development by interpreting complex planning controls and automating development assessment workflows.  

Prof. Thielscher’s role as a knowledge engineer is to extract expert knowledge from Professor Chris Pettit, Director of the City Futures Research Centre, and his team in such a way that it can be built into the AI system.  

“The team at CFRC approached me because they wanted to make their system smarter,” Prof. Thielscher said. “The plan is that eventually everyday Australians who are looking for a property can use Urban Copilot.  

“We have developed different personas or categories of users, and the system now reacts differently depending on whether the user is a developer or a member of the public, for example. It knows who it’s talking to and can tailor the answer to a particular profile.  

“We also use the engineered expert knowledge to verify the answers the digital assistant generates before providing them to the user, to make the system more reliable,” he said.

Digital skilled companions with Art & Design

Prof. Thielscher is also working with Professor Jill Bennett and colleagues from the School of Art & Design as AI lead on creating digital skilled companions offering psychosocial support.

“With their Art & Design experience, Jill’s team have built animated life-sized humans on large screens that you can have conversations with.

“These companions have particular life skills, for instance, dealing with issues like loneliness, and we are currently trialling these animated humans in an aged care home here in Sydney.”

In a previous trial, Prof. Thielscher said they “had some amazing experiences where people opened up and had conversations with the life-sized AI characters lasting for longer than an hour”.

“We want to make the systems smarter and more proactive. An important requirement is that we don't just provide expert knowledge, like a psychotherapist for example, but we also add knowledge from ‘lived experience’.  

“That is, we interview real people with issues such as an early onset dementia – one of the AI characters is based on a person with lived experience of dementia and the resulting AI character is very happy to share her experiences of that ailment.”

Prof. Thielscher said this lived experience makes the AI characters – who all have names – relatable, so that others who have similar experiences start opening up and asking about how the AI character dealt with the problem.  

“They get advice from the AI characters based on both the psychosocial support expert knowledge that I engineered and input, but also and very importantly, this bottom-up lived experience data. It also needs to be engineered so that it’s fed into the AI systems so that they talk and feel like real humans with that problem.”

AI animated humans that have a background story and expert knowledge are capable of going way beyond what current chatbots do, according to Prof. Thielscher.  

“Chatbots don't really have a long-term strategy. They don't do planning. They basically always agree with everything you say, and this isn’t what a skilled companion should do.  

“A good friend would occasionally challenge your beliefs, especially if you have a problem like loneliness. It's not a good idea to always agree – it's much better to gently challenge. And this is where the skills then come into play with our animated AIs.”

Prof. Thielscher and Prof. Bennett also have their own startup – PsychoSocialTech, skilled AI companions for loneliness.

“Our first customer was the Prince of Wales Hospital which wanted one of the characters for training of their own staff. So we built a character with an ailment like dementia – he lies in bed and the nurses use him for training on how to deal with patients like him.”

The startup has received grants for further research and is starting to see interest from countries like Japan and China. Japan, in particular, is interested because of its ageing population and their willingness to embrace new technology, Prof. Thielscher said.

Engineering knowledge or ‘planning’

This idea of making AI systems smarter by engineering knowledge, and in particular to show goal-oriented behaviour, is also known as ‘planning’ in Prof. Thielscher’s field. He has been working in this area since doing his PhD in the late ‘90s.

“We didn't have chatbots, like with these two ADA projects. I previously applied engineering knowledge and planning to build smarter systems, like robots. For example, if they want to demonstrate goal-oriented behaviour, they require underlying expert knowledge and they need to do some form of planning,” he said.

Another field he has been working on since 2005 is building AI systems that can play any game that they haven't seen before, just by being told what the rules are. Human programmers didn't provide any strategies or any further information.

“It was quite amazing because these self-learning systems turned out to be better players than human programmed players. This is what is known as ‘general game playing’ – you build systems that you only explain the rules to and then they learn all by themselves how to play a game very well.”

Prof. Thielscher said it’s very rewarding working on these two ADA projects.

“I have always found it fascinating to see what we can do with our theories in areas where it really makes a difference.”

Main image: This image is AI generated and is used in the context of MapAI.