AI and the ancient game of Go give new insight into expertise
27 July 2012
Using a traditional Chinese board game and artificial intelligence, researchers at the University of Sydney and Charles Sturt University have gained new insight into how expertise develops.
The findings, published this month in Nature's scientific reports (PDF, 1.5MB), will improve our understanding of how we think and help to develop more flexible artificial intelligences.
"In a rare achievement we used artificial neural networks, made up of hundreds of thousands of neurons each, to model how an expert rapidly evaluates a situation and narrows their choices down to the best options," said lead author Dr Michael Harré from the University's School of Psychology.
"As a species we are specialists, we can become experts in the most remarkably abstract tasks, but it has proven to be incredibly difficult to reproduce this because we understand it so poorly. This research has taken a significant step in our understanding by replicating the unconscious mental processes of experts in an artificial neural network and applying it to one of the most complex games we play today."
The researchers used thousands of records of professional and amateur matches of Go, a game for two players which originated in China over 2000 years ago.
"Using the data from these matches we replayed the amateur and professional games using our artificial neural networks," said Dr Harré.
"What we were able to do is model the mental processes that experts develop by using simplified versions of biological networks. Critically the networks we modelled not only change the way players think about the game, but they can literally change the way players unconsciously 'see' the game."
This is the first time that these subtle changes in how experts perceive their environment have been modelled. They are critically important for experts to recognise and use but are overlooked by non-experts.
"Importantly and impressively it is all done unconsciously using 'templates'. This refers to the library of patterns an expert builds to swiftly and efficiently cross-match the information they are receiving to identify what is important - before they have any conscious awareness that they are making those decisions, let alone how they made them."
The presence of the templates, what we could call 'cheat sheets' that the mind automatically uses to massively speed up its calculating, helps account for the speed and accuracy an expert can apply to making a decision. They are built up over a long period of time - 10,000 hours or more.
We might think of expertise as something done by air traffic controllers, neurosurgeons or Go champions, but all humans demonstrate expertise. Our ability to master speech, instantly recognise a face in the crowd or read the social cues in a conversation are all examples of very high-order 'expertise' that any artificial intelligence is still decades away from achieving.
"For this reason this research gives us new insight into the development of human thought. How do we develop our earliest perceptions into highly complex thinking?" said Dr Harré.
"It also promises to help develop artificial intelligence systems that are similarly flexible to humans; able to adapt to new learning environments while maintaining a stable and consistent 'mind'."
The research has been funded by the Australian Research Council and the US Air Force.
|Follow University of Sydney Media on Twitter|
Media enquiries: Verity Leatherdale, 02 9351 4312, 0403 067 342, firstname.lastname@example.org