OpenAI announced today that they’ve developed a new bot based on a neural network that can defeat human opponents in a game of Dota 2. While plenty of press has been given to neural networks that can defeat humans at a variety of abstract strategy games like Weiqi, this is the first time that an AI can genuinely hold its own in a complex video game designed with a human user interface.
Representatives of the lab were hoping to enter the bots in a series of matches at a championship event in August to play against the pros. This is in spite of the fact that AI players have traditionally had a difficult time mastering games like Dota 2.
Players have to make a huge amount of decisions at any given moment. A proper game of Weiqi can end in as few as 150 moves. OpenAI had to design bots that could make over 20,000 moves over the course of a single 45 minute Dota 2 game.
Engineers demonstrated last year that one of the bots could go up against a single human professional playing an abbreviated version of the game. They admitted, however, that they weren’t sure they’d ever be able to scale the AI up to a play a proper five-on-five match.
Researchers state that they figured out that an AI can explore and learn if given enough time to do so, and they configured a self-play environment for bots to learn in as a result. Two AI players in the system will go up against one another and learn from each other’s failures. Eventually, they were able to learn from what equates to around 180 years of play for every day they trained.
This was accomplished by using a massive stack of 256 different GPUs that featured 128,000 cores. While few gamers could ever hope to assemble a rig like this, OpenAI probably couldn’t have done it with anything less powerful.
Some critical gamers said that this isn’t truly an example of being intelligent because genuine intelligence is self-emerging and can learn without training. Training only polishes the skill of a human player.
Then again, regardless of how the bot players learned they’re certainly tough opponents for organics who have both innate skills and training under their belts.