Bots Compete To Be Most Human In Unreal Tournament
Writing by Evan Ackerman on Wednesday, 28 of January , 2009 at 12:57 am
Last year we met Elbot, winner of the 2008 Loebner Prize for successfully convincing three of twelve humans into thinking it was human by chatting with them. Chatting isn’t the only way to test the humanness of software, though, and 2K Games Australia decided to see if bots playing Unreal Tournament could fool human players into thinking that they were getting pwned (or not) by other humans.
The test was simple: a judge played a game of Unreal Tournament 2K over a network versus two other players, which were either moderately skilled humans or bots or one of each. After each game, the judge was asked which opponents were human, and which weren’t. There were some rules going in, like bots weren’t told what maps they were on, where weapons and stuff were located, or even the characteristics of each weapon (the weapons were modified for these matches). The bots had to learn the map, and learn which weapons to use when:
Some of these changes are made to require bots to learn something about the game, as a human can do, rather than simply following predefined scripts. Of course, it’s not necessary for this learning to be perfect! The competition is about making the bot appear human, not about making the bot play perfectly.
In the end, the most successful bot was AMIS, from Charles University in Prague, which was able to fool 2 out of 5 judges into thinking it was a human player. It had a “humanness” rating of 2.4 out of 5, while the least human human was rated 2.6.
Oh, and in the video above, 959 was a human while 228 was a bot. Were you fooled?
[ Botprize.org ] VIA [ KurzweilAI ]
Comments (1)
Category: Artificial Intelligence
- Add this post to
- Del.icio.us -
- Digg
Comment by Ironman
Made Wednesday, 28 of January , 2009 at 8:42 am
the charge through the line of fire was kind of an interesting move by 228 that really made it stick out.