Facebook will use Minecraft to give its AI a crash course in multi-tasking and listening to players’ directions
- Facebook is training an AI assistant how to multi-task using the game Minecraft
- Players are able to give the assistant commands like ‘build a circle’
- The bot will then read the command and execute the players request
- It will also learn from requests and develop new abilities using machine learning
Minecraft is more than just a game, according to Facebook, it also makes an ideal training ground for sophisticated artificial intelligence.
The company says it will use Minecraft, a popular sandbox-style building and adventure game, to train an AI assistant on one of the most important human abilities: multi-tasking.
‘Instead of superhuman performance on a single difficult task, we are interested in competency across a large number of simpler tasks, specified (perhaps poorly) by humans,’ researchers say in a proposal paper.
The bot will be able to interact with players and then perform tasks based on requests like ‘come here’ or ‘can you build a circle there?’
WHAT IS MINECRAFT?
Minecraft was created in 2009. At the start of the game, a player is put into a ‘virtually infinite game world.’
They can then walk around different terrains, including mountains, forests and caves.
Players can also fly up in the air for a birds-eye view of the landscape.
Players are given blocks and tools to build towns and cities.
As a player progresses they can earn advanced tools and building blocks in different materials.
The game was initially made for the PC but there are now Xbox 360 and mobile versions available.
Players can now enter virtual reality, with the game launching for the Oculus Rift.
For a human – even a child – those might be straightforward asks, but they can be monumentally difficult for current AI.
For instance, telling an AI to build a circle requires the technology to read the text instructions typed to it, know where to go, know what a circle is, where to place the blocks, how many to use, and much, much, more.
That difficulty, Facebook researchers say is part of the appeal.
‘In such a setting, understanding the content of a task can already be a challenge,’ reads the proposal.
‘Beyond this, a large number of tasks means that many will have been seen only a few times, or even never, requiring sample efficiency and flexibility.’
Ideally, the AI will learn from these requests and start to develop newer abilities to store in its repertoire.
According to them, Minecraft makes an ideal setting to train bots for several reasons.
One of its strengths as a training ground is that the game is extremely open-ended by design, offering a world of potential possibilities that the assistant can chew up and spit out.
Another ideal attribute, they say, is the fact that the AI is designed to be fun and as a result, players are actively incentived to interact with it.
‘Since we work in a game environment, players may enjoy interacting with the assistants as they are developed, yielding a rich resource for human-in-the-loop research,’ they write.
Minecraft, though simple to play, can be extremely elaborate. Picture above is a full-scale building created using the games modular building style.
The seemingly unlimited possibilities in Minecraft make the game an ideal training ground for an AI assistant, say researchers.
Even if the AI assistant fails in helping players build an item, it’s not hard to see how Facebook could use the loads of data from the bot’s foibles to inform future creations.
Facebook has long relied on algorithms to help make its platform more engaging, to weed out content, and offer up ad suggestions to its users — an ability that has helped drive its fundamental businessmen.
According to Facebook, an early version of the assistant can be downloaded here.