Microsoft’s pig wrangling competition is teaching AI’s to cooperate

I live in the countryside in the UK, and I also visited France a lot when I was a child, so I know just how difficult it is to wrangle a pig, and apparently it’s no different in the virtual world… who knew!? But as all us fine pig wrangler types know though it’s much easier to catch your pigs if you’ve got a friend, and that’s the focus of Microsoft’s latest Artificial Intelligence (AI) competition.

I bet you never ever thought you would hear the words “Microsoft” and “Pig Wrangling” in the same sentence – that is unless of course you’ve used their software and found it as difficult to use or implement as pig wrangling. And no I don’t want to hear about your implementation stories – sorry, this is my happy space.

Anyway back to wrangling Microsoft… er, pigs. It turns out that if you want to teach AI’s how to cooperate then virtual pig wrangling is a great way to go and Microsoft’s competition addresses potentially one of the biggest challenges we, and AI could face – how do you get a bunch of disparate, “individual” AI’s to cooperate with each other to achieve a particular goal?

We're goin' pig wranlin' Ma' 

AI researchers often develop software that’s capable of performing specific tasks and then measure how well it achieves it - whether it’s, for example, learning how to automate cybersecurity or law, code, create, and design new AI's, recognise images and sounds, run companies and replace CEO's, or just, simply, translate hundreds of languages. And that's the tip of the AI ice berg...

That said though there are very few tasks, certainly in the human world, where communication, social intelligence, and theory of mind, or the ability to anticipate and interpret another person’s actions aren’t valuable skills, but today very few, arguably none of today’s AI’s, possess any of these skills, and that’s what Microsoft wants to address. The competition could also open up new avenues of research that could one day help AI’s and humans work better together to achieve more than the sum of their parts.

“This is part of a broader trend of rethinking AI as ‘augmented intelligence’ rather than artificial intelligence,” says Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence.

During the Microsoft contest AI agents worked together inside “Project Malmo,” a special version of Minecraft, and Microsoft’s researchers designed the environment to make it easy to import and test different AI techniques and this is where the virtual pigs came into play – the AI agents in the game could either choose to work on their own to catch the pigs, or enlist the help of others, earning points each time.

Suffice to say the top teams in the Malmo Collaborative AI Challenge used cutting edge machine learning techniques such as deep learning to train their agents to work together which entailed feeding them large amounts of data. But some participants also made use of older, less fashionable approaches, something that seems to be coming back in vogue, that involved giving a virtual agent hard coded knowledge and understanding.

The winners of the contest, a team from the University of Oxford in the UK – yay, go UK - used reinforcement learning, which is a kind of machine learning inspired by the way animals learn via experimentation. Their agents experienced positive reinforcement whenever they successfully worked together to grab the pig.

“There was no single type of approach that emerged as a clear winner,” said Katja Hofmann, the lead researcher on the project, adding, “it’s likely that hybrid approaches will prove particularly promising directions for future research.”

The pig wrangling challenge takes inspiration from a thought experiment known as the Stag Hunt, which explores game theory concepts, a branch of mathematics concerned with cooperation and negotiation strategies, where the idea is that two hunters must decide whether to hunt a hare on their own or team up to snag the bigger prize of a stag.

The top teams involved in the contest, judged according to the score they achieved as well as the novelty of their work, will receive a $20,000 research grant and a place at Microsoft’s Research AI Summer School.

Pedro Domingos, a professor at the University of Washington who studies machine learning and data mining though says training AI software inside simulated environments has its drawbacks.

“Software can become over optimised for that particular environment and therefore less useful in the real world,” he says, “although more sophisticated simulated worlds are starting to change this.”

Domingos adds that cooperation between humans is so complex and subtle that it is hard to imagine the Microsoft project producing genuinely useful approaches, but, despite some scepticism, he’s encouraged by it.

“It’s still early days in this area, and Minecraft is an environment with a lot of possibilities,” Domingos says, “[it’s] richer than things that have been used before, so it certainly seems worth trying.”

While Microsoft’s competition is, at the moment, amusing, it obviously has the more serious objective of trying to teach AI’s to cooperate with each other and, depending on your point of view, that could be a great thing, or a bad thing. On the one hand imagine AI’s cooperating with each other to solve Cancer, or at the other end of the spectrum cooperating with each other to take down an air traffic control network. All of a sudden you realise the slippery slope we find ourselves on…

Map into that the fact that other experiments are teaching AI’s - which lest we forget are still regarded as black boxes - how to negotiate, not just with each other but also with people, and compete against, or depending on your point of view, fight each other, and we’re starting to enter a very uncertain new world. And all of this is happening when we still don’t have a proven way, despite work in the area, to either query their inner workings, or turn them off with kill switches.

Compared to what we’re going to have to face in the near term future, all of a sudden pig wrangling looks simple. Here piggy piggy piggy…