In contrast to AlphaGo and AlphaGo Master, AlphaGo Zero did not have a hardcoded subroutine to help evaluate ladders in go 囲碁. Did Zero learn to evaluate ladders correctly? If so, examine its neural network to try to understand how it does it. Ladders which turn corners can require reading a hundred or more moves ahead to know whether the ladder works.
How well do AlphaGo and successors play combinatorial endgames as made famous by the book Mathematical Go?
No comments :
Post a Comment