First they conquered StarCraft II, now Dota 2; it’s almost as if these AI research companies want the machines to win in the coming robotpocalypse.
A competition was recently staged between professional, human, Dota 2 champions, team OG, and a team of bots, OpenAI Five, as a test of OpenAI’s skill at training AI. According to The Verge, the bots steamrolled the humans using aggressive techniques that favoured short-term gains, rather than long-term payoff, such as using in-game currency to insta-revive dead heroes early on; a tactic usually reserved for emergencies.
We see this happen in test games all the time: the bots buy back, the humans laugh, and then the humans lose. Hard to know if it’ll happen here too…
OpenAI trains its bots using reinforcement techniques; dropping neophyte AI into the game and allowing to learn by trial and error, using rewards and incentives to keep it interested in progressing. According to Greg Brockman, OpenAI’s co-founder and chairman, “OpenAI Five is powered by deep reinforce learning, which means we didn’t code it how to play. We coded it how to learn”. The bots apparently experienced the equivalent of 45,000 years of Dota 2 gameplay in just 10 months, so maybe OG shouldn’t feel too bad about losing.
So, the robots just aren’t very good at long-term planning, apparently, but does the long game even matter if you never get there? Since they have no time for drawn-out strategy, at least we know we won’t have to suffer long.