News for hackers. Stuff that matters.
Re:OpenAI Five by backpropaganda on Monday June 25, 2018 @02:56PM (#17394762)
> Dota is far too complex for random search

Why? We know that random search is smart enough to find a solution if given arbitrarily large computation. So, that random search is not smart enough for Dota with the computational budget you used, is not obvious. Maybe random search would work with 2x your resources? Maybe something slightly smarter than random search (simulated annealing) would work with 2x your resources?

> and if that weren't true, it would say something about human capability

No it would not. A human learning a game by playing a few thousand games is a very different problem than a bot using random search over billions of games. The policy space remains large, and the human is not doing a dumb search, because the human does not have billions of games to work with.

> See our gameplay reel for an example of some of the combos that our system learns

> Our system learns to generalize behaviors in a sophisticated way.

You're underestimating random search. It's ironic, because you guys did the ES paper.

There are no comments on this comment yet.