Deepmind - a leading researcher into artificial intelligence and its potential for positive use - has just announced a partnership with Blizzard Entertainment, using classic strategy game StarCraft 2 as a basis for improved AI learning.
The pair have created a set of tools - known as SC2LE - to accelerate AI research specifically within StarCraft 2. In addition to a machine-learning API to allow both researchers and developers access to the game, Blizzard is also providing a dataset of anonymised game replays - currently at 65,000, this will increase to 'more than half a million' in the next few weeks.
In a post on its website, Deepmind explains that part of the draw of StarCraft 2 is the multi-layered gameplay and the prevalence of 'sub-goals' within the game - for example, gathering resources whilst also engaging in combat with the enemy. This also means that often players have to play the 'long game', putting into place actions which may not play off until much later in the game.
Although the AI 'agents' are able to perform tasks well in isolated mini-games, when exposed to the full game they struggle with tasks that humans find fairly trivial - it's this that Deepmind hopes to combat with Blizzard's help. The most efficient way to train AI is through 'imitation learning' - learning from the human examples provided, and the sheer number and variety of replays that Blizzard is able to provide from StarCraft's huge player-base gives them plenty to work with. The video above shows the difference between a 'trained' and 'untrained' AI agent.
Although it may not seem like a huge scientific leap to teach AI how to play a video game, in reality the skills that are being learned - sequence prediction and long-term memory - are hugely important to the future of true AI.
Let's just hope they don't make their own language.
Matt Davidson is a freelance writer for IGN who is equally excited and frightened by the prospect of AI. Follow him on Twitter.
Aucun commentaire:
Enregistrer un commentaire