playing passes instead of good moves

Issue #38 resolved
dsmic created an issue

The passes are played, even if the winrate of other moves might be better. Reason seems to be the RAVE value of PASS. I attached a sgf playout, and one can see, that the rave value i 5/5, which is the rave_init_wins. This leeds to overestimating of the value of pass if the number of playouts is low?!

In this case you get PASS: Wins/Playouts: 1621.3/6091(0.26618) RAVE Wins/Playouts: 5/5(1)

D12: Wins/Playouts: 641.788/1742(0.36842) RAVE Wins/Playouts: 1984/5709(0.347521)

Comments (3)

  1. Francois van Niekerk repo owner

    It is a little weird that this happens. If there are 1000 normal playouts versus 5 rave playouts, the RAVE formula should correctly weight the rave value to be less than 1% of the total value.

    Perhaps weighting the rave value towards 0.5 instead of 1.0 would be a better "fix"

  2. Log in to comment