Using a similar technique to Google when it taught its Deepmind network to play Atari games, two researchers from Carnegie Mellon university have created a self-taught Doom bot which became the most efficient killer at a recent competition. Known affectionately as Arnold, the bot placed second in the end, though it walked away with the best kill to death ratio.[yframe url=’http://www.youtube.com/watch?v=NgndlFybEho’]
The difference between bots at the Visual Doom AI Competition (thanks Kotaku) and say your average cheating bot in other games, is that these ones were only able to use the information on the screen in front of them and their memory of recent screens. Essentially, they had to act like real players. That’s what makes Arnold so impressive, because it behaves like a really efficient human player.
Admittedly so do most of the bots on show during this contest, though some of them do appear to get stuck on the odd wall, or spin in place as they try and get a bead on an enemy. In reality it still seems likely that most of these bots would be wrecked by an efficient human player.[yframe url=’http://www.youtube.com/watch?v=94EPSjQH38Y’]
But they are getting smarter and that’s the point. These bots can continue to get better forever, they have the time to play Doom all day, every day. Their different memory functions are able to keep them learning from their mistakes and remembering what they do scene to scene so they can hunt down the opposition that much more efficiently.
Arnold suddenly seems like quite a fitting name for the most efficient killing machine.
Discuss on our Facebook page, HERE.
KitGuru Says: This is the kind of AI that you want people developing, as it at least puts you on an even playing field with them. If they can wall hack and have 100 per cent accuracy, then what’s the point?