“We estimate the probability of winning to be above 95%”, said OpenAI’s Dota 2 bots at the start of their second game against extremely skilled humans. I knew they’d been trained up using a sophisticated reinforcement learning technique that had instilled them with millenia worth of experience. I didn’t know they’d been versed in trash talk.
It seems they had training time to spare, because they handily won two out of three games at the “OpenAI Five Benchmark” yesterday. This was the robo-team’s first outing against non-amateur players, in preparation for the show match they’ll play at the International 2018. There, they’ll face off against the best players in the world. The bots won their first game here in 14 minutes, so I don’t fancy humanity’s chances.
The bots were up against a team of ex-pros and analysts, all ranked in the top 99.95th percentile of players. The humans had the likes of ex-pros William ‘Blitz’ Lee and Ben ‘Merlini’ Wu in their ranks, but were stomped regardless – until game three, when the OpenAI team agreed to throw the bots’ hero selection over to Twitch chat. They still managed to hang on for half an hour, but the bots knew they’d been put in a tough spot and only gave themselves a 2% chance at victory.
It’s worth noting, however, that they weren’t playing normal Dota. Although some of the more severe restrictions were lifted in the weeks before this event, the bots were still paddling in a pool of 18 available heroes rather then Dota’s full compliment of 115.
I dug into the details with Mike Cook a couple of weeks ago, an AI researcher known for his game-making AI Angelina. As he predicted, the bots fell apart the moment they were pushed into a situation they hadn’t practised for. Based on what I’ve seen here, the bots seem to be exceptionally good at ‘pushing’ line-ups, where they’ll aggressively march down lanes using abilities that excel at taking towers. When those hero picks were taken away from them, their decision making was thrown out of whack.
I haven’t watched the matches in full yet, but I did catch a snippet of Kevin ‘Purge’ Godec – ex-pro, analyst, coach and commentator – remarking on exactly the same thing. It’s neat that Cook and Purge saw the same limitations from their different perspectives.
You can watch all of the games below, as well as listen to OpenAI explain the machinations behind their machines.