SoylentNews is people
SoylentNews
Submitted via IRC for BoyceMagooglyMonkey
AI agents continue to rack up wins in the video game world. Last week, OpenAI's bots were playing Dota 2; this week, it's Quake III, with a team of researchers from Google's DeepMind subsidiary successfully training agents that can beat humans at a game of capture the flag.
As we've seen with previous examples of AI playing video games, the challenge here is training an agent that can navigate a complex 3D environment with imperfect information. DeepMind's researchers used a method of AI training that's also becoming standard: reinforcement learning, which is basically training by trial and error at a huge scale.
Agents are given no instructions on how to play the game, but simply compete against themselves until they work out the strategies needed to win. Usually this means one version of the AI agent playing against an identical clone. DeepMind gave extra depth to this formula by training a whole cohort of 30 agents to introduce a "diversity" of play styles. How many games does it take to train an AI this way? Nearly half a million, each lasting five minutes.
Source: https://www.theverge.com/2018/7/4/17533898/deepmind-ai-agent-video-game-quake-iii-capture-the-flag
(Score: 2) by acid andy on Tuesday July 10 2018, @09:45AM (3 children)
Uhh, that sounds less like intelligence and more like good, old-fashioned, brute forcing to me. Who cares if it involves a neural net if you have to throw that much time at it?
Master of the science of the art of the science of art.
(Score: 2) by looorg on Tuesday July 10 2018, @11:26AM (1 child)
That is a fairly valid point. Also 450k games a 5 minutes each is 2.250.000 minutes or about four years and change. I wonder how good a player would be if they played that many minutes of the game, I'm not sure there are any players like that alive today. The previous numbers assume you are playing around the clock 24h per day everyday which no human can sustain. Even if one was to play for just 12h per day that would be 8 or 9 years of playing QuakeIII (or whichever game one picks) every single day for 12h non-stop.
I'm sure one could argue that humans learn in other ways all the time, life experience bleeds into the gaming experience etc such as we know how to open doors etc. But I still doubt it's comparable. Perhaps it goes back to that idea of how long it takes to master something; the previous idea of it taking 10.000 hours was deemed to be bullshit years ago, which would even if it was true only be about 600.000 minutes which is still a lot less then the FTW-AI got to play the game.
(Score: 2) by FatPhil on Tuesday July 10 2018, @03:19PM
It's 4 months and change for a small group of mates who then go on to nominate their champion - a totally feasible endeavor.
However, it's 4 hours and change for players of a reasonably popular online game who then go on to chose their champion. That's the better comparison for the work factor involved.
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2, Insightful) by Kalas on Tuesday July 10 2018, @09:52PM
Please consider that given a game using entirely AI players, there's no reason the devs can't just set the game to run 10 or 100 times faster than realtime.
Processing power is pretty much the only limitation for how fast these simulations can be ran. You could get all those thousands of 5 minute (subjective time) games done in weeks or maybe days given enough cores to throw at it.