Indeed, the GPU is very important. About half a year ago I let LC0 attend the HGM tourney a few times on my GTX1080ti and it did clearly better, if I remember well the last time it attended it won the tourney, after this I stopped with it.Modern Times wrote: ↑Sun Mar 17, 2019 11:32 amIndeed it did not do that well. In the absence of Arasan and Texel I thought it could fight for the win, but at the end of the day the £100 GTX 1050 was probably not enough. A £700 - £800 RTX 2080 would have been a different story, but I don't have that kind of cash to be throwing at a PC just so it can play Lc0 at a high level. It is also a horribly flawed engine in many ways, as we saw in one game when it threw away a +40 advantage to end up drawing. Very frustrating engine also, for example where instead of making a pawn move to promote, it just aimlessly decides to shuffle pieces around for a while instead. Or else it decides to throw away pieces because it doesn't need them to secure the victory. Until it solves all those issues, it is not a top engine in my book.Joost Buijs wrote: ↑Sun Mar 17, 2019 7:28 am I expected LeelaRB to win the tournament, but somehow it didn't live up to expectations.
A neural network is very capable at encoding positional information, but tactics and endgame play where you need precise and deep calculation is not it's strongest point. Aimlessly shuffling pieces around during the endgame is not a neural network problem as such, but more an error in the way it is trained. Like HGM already said, you also have to encode depth to win in the network, otherwise it doesn't know how to make progress. When you use a table-base with only win/loss/draw information and without knowing depth to win or depth to conversion, you will encounter exactly the same problem.