TalkChess.com

[Antonio Torrecillas]

Hi Vincent,
How strong is that stuff?
Rocinante 1.01 PB* with ufo evaluation (material+pst) 1650 elo
Rocinante 2.0 PB* with ufo eval 1750-1800 elo
Rocinante 2.0 PB* with new eval not tested.
Rocinante 2.0 MCTS AB with ufo eval 1850-1900 elo
Rocinante 2.0 MCTS AB with new eval 2000 elo.

Why MCTS AB?

The main drawback of PB* is a double evaluation for leaf nodes (real + pessimistic). The second is a kind of horizont effect, and a last chance effect. If you expand a node and find it quiet you never come back to this node to see deeper.

I tried to overcome the first drawback, doing simply a single leaf evaluation. To overcome the second I was trying to change the way PB* trace down the next node to expand.

Once there I realized that I do reinvent something that already exist. So I applied Upper confidence to tree and a single eval.

So that's a MCTS with the random playout replaced by a fixed depth alpha-beta.

I agree that MCTS with random playout is really weak, but a UCT on top an alpha-beta is another beast.

This is a newborn, let see how it grows.

regard Antonio.