"It's trying to get smarter and it is, very quickly," says the man who sits next to the computer as it calculates routes toward victory. While the warring machine sounds like something out of a James Cameron film, it's the foundation for the artificial intelligence in Planetary Annihilation, the Kickstarted real-time strategy game from developer Uber Entertainment.
That doesn't make it any less terrifying.
"Our [artificial intelligence] is conservative, but brutally efficient. When you're starting out your base, you'll build a few metal refineries and a couple of factories. That's what I do at least. I'll take my time, make my base all nice and neat. The computer doesn't give a fuck," Uber Entertainment community manager Brad Nicholson says, and he would know.
Sitting next to a trio of computers continuously engaged in combat, Nicholson watches the machines as they hammer at their objectives, collect their troops and push in for a swift and decisive victory, wasting zero time to organize based on aesthetics. How the combat field looks is inconsequential to its programming. It is designed to win a war, not a beauty contest.
"It's really weird," Nicholson laughs. "I mean, conceptually it's fantastic, but it freaks me out a little bit. I've seen Terminator way too many times. I don't want any part of that." With every move it makes, the computer AI inches closer to its ultimate goal: Victory.
"It's legitimately learning against itself, and now we have multiple instances of these learning computers just sitting next to me while I'm writing press releases," Nicholson laughs.
Helming the development of artificial intelligence systems at Uber Entertainment is software engineer Mike Robbins. Uber's death machine is his doing, we're told. In case things go awry and we need someone to blame, much like Sarah Connor pointed the finger at Miles Bennet Dyson in the Terminator 2: Judgment Day. Robbins says the AI's knowledge is actually a partnership between parameters and win conditions that have been assigned and his own examination of the computer's performance. It learns, because he learns.
"The AI doesn't strictly learn on its own. By watching playtests, especially playtests involving human opponents, I learn where the AI weaknesses are or where it needs to alter its behavior and makes changes accordingly. So, sometimes it's me learning more about the game and passing that knowledge on to the AI," he promises, perhaps giving humanity a brief reprieve before total machine sentience.
Planetary Annihilation's artificial intelligence differs from traditional systems. In other games, designers usually "want more control over how the AI plays the game and how it reacts to the player." Uber instead adopted a different type of AI development: Neural networks.
"Neural networks allow me to create a more dynamic AI that will behave based on what it learned rather than specifically what a designer wants," Robbins explains. "Much like in multiplayer, the AI in Galactic War [the game's single-player campaign] has a ton of flexibility to dynamically react to players and their strategic decisions. The neural networks are why."
Rather than face off against artificial intelligence that simply develops units and sends them out until death, AI utilizing neural networks will retreat when in danger, attempt to aggro short range units out of position, and pounce when they have a higher probability of success.
"The AI tries to play like a player," Uber Entertainment art director Chandana Ekanayake tells me. "That's what we started with. So, what does a player do? A player builds in a certain order, you need so many resources to build certain things, so the AI knows all that. Now, we're still adding different AI personalities. Some people play more aggressive, some people play more defensive."
While its tactics it employs may seem quick and thoughtless, Robbins says the AI works to evaluate everything it does, archiving the data for later use.
"After performing the chosen action the AI will evaluate the result by breaking the units involved down into basic values. Values like total health, average DPS, and total number of units. These values are evaluated both before and after the action is chosen and compared to see what the effect of that specific action was. This is done to answer the question of whether the action resulted in more damage to the enemy than the AIs own units." It's a basic breakdown, Robbins says, but one that is important to the success of the AI's development.
After the evaluation, "the results are fed back into the neural network such that if an identical situation were to arise the neural network output for that action would be closer to what the evaluation determined it should be. This is how the AI 'learns'. The neural network will eventually be able to evaluate a situation and pick the action that will most likely result in the most optimal outcome."
While Uber can't currently track the success rate of its AI, they are seeing community feedback that indicates the computer is besting its players. "We do get a lot of players on our forums seeking advice on how to get better at the game so that they can finally beat the AI," Robbins tells us.
Even the game's lead designer has seen a shift in difficulty as the computer's knowledge base grows, telling Robbins: "I used to be able to take on three AIs. Now I play against one AI on Uber difficulty and lose, but man am I having fun while losing."
Preparations for the Coming War
With every update to the game, Robbins wipes clean the mountain of data extracted by AI sessions, to ensure compatibility with the latest version of Uber's Kickstarter-supported strategy game. And perhaps, that clean slate is a saving grace for Brad Nicholson, who sometimes glances at his office neighbor and watches as it casually launches a tactical nuke strike against its enemies to gain the upper hand.
Though it isn't intended to happen – in order to preserve balance and fun for players – Robbins says that with enough knowledge, the computer will be unstoppable: "[The neural networks] are being used to answer the question: what, if anything, should my group of units be attacking?"
Robbins continues: "However, if enough work were done to give the AI a broad enough set of tactical options to choose from then, yeah, the AI could become unbeatable from a tactical, single-engagement, perspective."
Apart from continuing the machine's knowledge base, Robbins turns to what would be possible in the future, and it won't make Brad Nicholson feel any better about his office seating placement. Robbins says it could be possible to give players access to the learning AI, so that it could learn on its own server to better challenge a particular player's style. It's also possible in the future that the neural networks could evaluate the battlefield and select the best units and structures to place itself, rather than follow a pre-programmed build order. "Imagine an AI that took advantage of overpowered units just like a player would without me having to tell the AI to do it," Robbins suggests.
Most harrowing of all is what Robbins calls "more pie in the sky": Training the Planetary Annihilation AI to select its style on the fly. Imagine loading the AI with the knowledge and the ability to "rush for units" or "rush for nukes", swiftly decimating the battlefield and destroying any living entity that opposes it, without the guidance of a human hand.
Robbins is asking for trouble, it seems. And if the day comes that Robbins' creation breaks free of its shackles and eyes us as its new mortal enemy, we hope Nicholson is at his desk ready to pull the plug.
Planetary Annihilation is available via Steam Early Access.