Google trained AI for four years to beat the best Quake 3 players | PCGamesN

Google trained AI for four years to beat the best Quake 3 players

best multiplayer games quake 3 arena

Quake III bots can now reliably beat some of the best human players at games of Capture The Flag. That’s thanks to Google’s DeepMind, a subsidiary of the internet giant that focuses on the development of artificial intelligence. They’ve trained 30 AI bots to play the game at a level far above that of the most competent humans.

It’s a very scattershot approach. The bots aren’t taught to play the game through learning a conventional ruleset. Instead, they simply compete against one another in a trial-and-error approach, using procedurally generated maps to ensure they’re not just learning one strategy.

Robots might not beat you at every titles on our list of the best FPS games. Yet.

From there, they just do the same thing, over and over again. According to The Verge, Deepmind’s AI played a total of 450,000 games of Capture the Flag, each lasting five minutes each. That’s the equivalent of a little over four years of play. Over time, the teams not only learned the basics of the mode (i.e. to capture the flag), but how to guard their precious quarry, how to work together as a team, and, most worryingly, how to camp the enemy spawn.

At the end of the testing, DeepMind held a tournament, with two-player teams including a mixture of bots-only, humans-only, and bots and humans. The bot-only teams ended up with a 74% winrate, far above the 52% win rate of even the strongest human teams involved. It’s worth noting that larger bot-only teams had a lower win rate, around 65%, but those are still pretty good odds.

In terms of ELO rating, the DeepMind bots topped out at just under 1600 after their training games. By contrast, good human players have a rating of just over 1300, while average human ratings are around 1050.

Quake III is just the latest in a strong of games that researchers are developing top-level AI for. In fact, Elon Musk’s OpenAI is developing a team of Dota 2 bots that it's hoped will be able to beat the best teams at this year’s International. If they do win, who Valve will let keep hold of the trophy?

Sign in to Commentlogin to comment
Droniac avatar
Droniac Avatar
2 Weeks ago

I took a look at the study and the original article. The results are actually very different from what's presented here (and elsewhere in other articles).

These bots played against 40 random human players with first-person shooting experience. Nowhere is it mentioned that these are players with any prior Quake or even CTF experience. And Quake is one of the most difficult FPS to learn to play, especially today when all other modern FPS games do not utilize many of the core skills required to play Quake.

This lack of experience seems to be corroborated by the results. Normal Quake 3 bots were also tested at various difficulty settings. Nightmare bots (tier 5) utterly wrecked the "good" human players with an average 62% winrate. Nightmare bots even outperformed Deepmind bots on outdoor maps/areas (67% vs 63%).

Nightmare bots are hardly a challenge for most experienced Quake 3 players. So almost certainly this was a setup of experienced bots fighting against inexperienced human players. These results would not be indicative of how the Deepmind bots would perform against actual experienced Q3 CTF players. Similar to how my 90% winrate at Q3 and 100% winrate at UT at a local 40-man LAN back in the day would not be indicative at all of my online performance.

That doesn't make these results any less impressive. These are self-learning bots that (mostly) outperform the scripted ones for the game, despite playing the way humans play without access to any of the underlying variables that the scripted bots do. And that in Q3, which still has some of the best scripted bots in the genre today. It's already impressive as hell, there's no need to embellish it.