Could AI be used to save lives in the real world?
On Sunday 18th March 2018 a 49 year old woman in Arizona was killed by an autonomous Uber car, which struck her as she pushed her bicycle along the roadside. The death was blamed on defective software. Two years prior to this, the first of multiple Tesla driver deaths occurred. There is significant evidence which points again to defective autopilot software. So far, not so good.
As a keen video game recruitment specialist and player, I wondered whether advances in AI which are being used to beat humans at video games could be transferred to vehicle software, and save lives in the process. After all, AI has shown a remarkable capability to exceed human ability in unexpected areas.
AI first showed its potential in 1997, when IBM’s Deep Blue AI won a game of chess, under tournament conditions, against the Russian chess grandmaster and then reigning world champion, Garry Kasparov. By 2015 Google’s AI subsidiary Deepmind had trained AI to play 49 different video games from an Atari 2600, beating a professional human gamer’s top score in 23 of them. Later that year Deepmind’s AlphaGo became the first AI to beat a professional Go player, before beating Go world champion Ke Jie in 2016. It learned to play Go – a complex Chinese board game dating back more than 2,500 years and generally considered more complex than chess – by learning from 30 million moves played in human-on-human games. In 2017 Deepmind developed a successor to AlphaGo, named AlphaGo Zero. Instead of relying on human teachings, the AI learned by playing games against itself. This new version went from not knowing how to play the game at all, to surpassing the abilities of the version which beat Ke Jie, all within just three days.
Methods of training AI are continually improving. This year, Deepmind revealed its new and innovative approach. Using their own DMLab:30 – a training set built on ID Software’s Quake III and an arcade learning environment running 57 Atari games – Deepmind has created what it calls Importance Weighted Actor-Learner Architectures (IMPALA). Deepmind believes IMPALA is ten times more effective than traditional methods of training AI. Rather than learning to play as a single player, IMPALA is capable of playing 30 games simultaneously using one ‘brain,’ and is able to achieve a throughput rate of 250,000 frames per second. Whilst not the fairest comparison, on a modest gaming system a human player is exposed to just 60 frames per second.
Just last month, bots developed by OpenAI – a non-profit AI research company championing safe AI –succeeded in defeating four pro players and one commentator in a best-of -three series of games of Valve Corporation’s MOBA game DOTA 2. The AI is self-taught, learning at an astonishing rate of 180 years per day.
In each of the above cases, Deep Blue, AlphaGo Zero, DMLab:30 and OpenAI have learned or are learning to play complex video games and beat professional players at an increasingly rapid rate.
If AI can teach itself to beat human world champions within just a matter of days, what’s to stop it developing life-saving interventions to enhance those already available through in-car telematics? After all, LexisNexis Risk, an information business which specialises in insurance information, points out that 94% of all car accidents are caused by human error. Research has consistently shown that experience is instrumental in improving driving behaviour. Good2Go, a US insurer, says that
According to both the NHTSA and the IIHS, the safest drivers are between 64 and 69 years old. Data reveals that male teenage drivers are the most dangerous drivers on the road. However, when adults reach their 80s, they become riskier drivers as their visual and cognitive skills begin to fade, causing them to make more traffic mistakes. In fact, drivers 80 or older are involved in 5.5 times as many fatal crashes than middle-age drivers.
Given that AI programs can learn good driving behaviour exponentially faster than human beings, and that they do not experience the loss of visual and cognitive skills that older drivers eventually
suffer, it seems only natural to assume that the technology will soon be utilised by OEMs and insurers. Indeed, as long ago as 2015, researchers at Cornell university were using AI to predict and enhance driver behaviour; and AI-enabled cameras are already being deployed in commercial vehicles to detect and understand problems with the driver.
The technology is out there, backed by some of the biggest names in tech and advancing at a mesmerising pace. I’ve always thought video games would save the world. Surely now it’s only a matter of time before they start to save human lives.
James Dodd
Martin Tripp Associates is a London-based executive search consultancy. We work across the media, information, communications and entertainment industries, including video game recruitment. This means we can bring best practice from across the sectors to your business. We have also worked with some of the world’s biggest brands on challenging senior positions. Feel free to contact us to discuss any of the issues raised in this blog.