Archive This article is from our archive and might not display correctly. Download PDF
DOTA 2 is probably the most complex, demanding, strategically deep game played by pro gamers today. Teams of five battle it out on a relatively large map to farm money, gain powerful items, and ultimately win by destroying the opposing team's base. The game takes hundreds of hours to learn properly, let alone master. Each game takes around forty minutes to complete, but can last over an hour, which means planning and teamwork are crucial to achieving success. It's like playing ten games of chess simultaneously, except you can't see half the board, and the people controlling the rest of your pieces are all idiots. All these factors make designing a robot capable of playing DOTA pretty hard. This summer, an artificial intelligence non-profit backed by Elon Musk thinks they can do it.
OpenAI's DOTA 2 bot made its debut last year, soundly stomping professional player Danylo 'Dendi' Ishutin in a 1v1. DOTA 2, however, is a team game, and Dendi's match last year represented a very small part of the title's gameplay. The artificial intelligence is a slow learner, and the OpenAI team have introduced it to small sections of the game, bit by bit, over the last year. In an extraordinary display of modern computing power, the bot crunches through the game at a highly accelerated rate, and plays 180 years of gameplay each day against itself.
OpenAI use 'reinforcement training' to teach the AI the difference between 'good' and 'bad' decisions. Useful objectives like kills are assigned positive values, and the bot works to raise those values. The bot's capabilities have improved vastly over the year. It began by learning just five heroes. Now it knows 18, and can play against opponents on varied characters too. The learning process is both extraordinarily impressive, and also a sad indication that machines still seem to learn much slower than people. A human takes around 13,000-20,000 hours to reach the professional level at DOTA. The AI has taken hundreds of lifetimes.
Is this slow learning process enough to face the best? OpenAI wants to find out by having their creation face five DOTA 2 players in August in San Francisco. Those players will each be in the 99.95th-percentile of DOTA ability and includes Merlini, Blitz, Cap, and Fogged, some of whom are former professionals. Gameplay limitations will still be in place, but will be significantly less strenuous than last year. Notably, the bots are now capable of handling wards: items that allow you to view more of the map simultaneously. The upcoming match will only place limitations on couriers (the entities that deliver items to the player,) and two items: divine rapier and bottle. Even with those limitations, success by the bot would be extremely impressive. Unlike chess, a key part of DOTA's metagame is teamwork, and amazingly, that's where its systems flourish. The OpenAI developers have even had to limit its reaction time to 200ms to more closely mimic humans. It's nice to know we now have to dumb games down to give people of flesh and blood a chance.
So, it turns out robots have no ego, and are actually at the stage where they require handicapping to be able to make fights fair. Video games might be an easier frontier to master than real life, but this AI's capability is an incredible (and a little concerning,) vision of the future. Completely autonomous drone warfare beating actual pilots? Success here could be exciting and terrifying in equal measure.