Hacker News Reader

The standard version of shogi.

I wonder if the extra space and pieces would increase the length of training polynomially or exponentially.

Go was considered to be difficult for computers because of the high branching factor compared to chess. DeepMind still beat the best humans there pretty comfortably, with similar amounts of training to what allowed it to compete with the best engines in chess and shogi.

I'd expect AlphaZero to hammer the best humans given a similar amount of training (~100,000 2017 TPU-hours) in this shogi variant, so no increased training time at all by that metric. Maybe a more meaningful metric is how closely it approaches perfect play, but we don't have a good way to measure that.

I am also more curious about the 9-dimensional Space shogi, https://en.wikipedia.org/wiki/Space_shogi ... "Black starts the game occupying levels 1 through 3; White starts on levels 9 through 7."

Isn't it just 3-dimensional (9x9x9)? That's a smaller playing space than taikyoku shogi, it's again a game equivalent to chess (in the von Neumann sense) and a computer wouldn't find anything more conceptually difficult about moving through a 3d array than a 2d one. So I'd expect the computer's advantage to be even larger here.

Reply on news.ycombinator.com