Artificial Intelligence Programs Kill Each Other In ‘DOOM’ For Our Amusement/Doom

We can all agree that teaching artificially-intelligent robots how to arm themselves with rocket launchers and create a murder carnival is just a fantastic idea. But these robots aren’t ready to be let loose outside our front door (yet). They need to train first. Luckily (luckily?) there’s VizDoom 2016, this year’s competition pitting AI programs against each other in a game of DOOM.

Why DOOM? So we’re ready if Hitler comes back as a robot, of course. Or perhaps because the influential 1993 game has simple 3D maps and multiple gameplay styles, which makes it a handy touchstone for AI research.

This year’s VizDoom held two competitions on Thursday. The first involved AIs with rocket launchers fighting each other to the death on a map they’d been given in advance. Facebook’s team won that competition in 10 out of 12 matches. The second, more complicated competition used an unfamiliar map with varied weapons scattered around, requiring the AI programs to find their own way and pick which weapons to use. Intel’s team won 10 out of 12 games by a wide margin, so you know which company to run to right now if the sh*t goes down.

All of the AI were built on the same platform, VizDoom, and are usually trained to play the same way us humans game: reinforcement learning. Quartz puts it into perspective:

Pretty much every modern video game claims to have AI, which is used as a blanket term for characters programmed into the game that seemingly act autonomously. However, they’re programmed to have certain behaviors, act within certain limits, and don’t learn.

The AI in the VizDoom competition don’t work that way. They play the same way any human would: by looking at the screen, processing what’s happening, figuring out the best strategy to win, and then actually controlling the character to achieve that goal.

One competitor to watch out for in next year’s VizDoom is the University of Essex’s program by Dino Ratcliffe. His AI is programmed to be a dirty spawn-camper, waiting at spawn points to kill every competitor just as they’re respawning. That computer broke the day before this year’s competition, so a replacement with only 40% of its training had to compete. It still took third place. Next year, expect a fully-trained version of this spawn-camping bastard to be a bot to watch out for.

Seriously, though, an AI program that spawn-camps would seem so human in a real-world game. But we refuse to regard it as truly intelligent until it learns how to properly pick its own initials…

(Via Quartz)