Game-AI Forum

Hi,

Has anyone ever tried adding a transposition table to an UCT/MCTS implementation?

I tried without much success. It works, but it plays worse than the non-tt version.

What I do is: for each position I have a node. This node stores pointers to children (of course as it is a tree) and to all the parents (can be multiple due to transpositions).
Now the score function is:
Here I use the sum of all the visit-counts of all parents. Is that how it should work? Or am I walking an uncultivated area?

Hi,

Using a transposition table is a good idea, and is rather common. It should make the program stronger.

You should not have to keep pointers to parents. If I understand what you do correctly, you may get very different "parent_visit_count" for every child of a node. This does not seem good.

Also note that UCT is not very efficient formula. Using a policy with the AlphaGo formula will be much better.

Rémi

Rémi Coulom wrote: ↑

Fri Apr 22, 2022 8:00 pm

Using a transposition table is a good idea, and is rather common. It should make the program stronger.

You should not have to keep pointers to parents. If I understand what you do correctly, you may get very different "parent_visit_count" for every child of a node. This does not seem good.

Ok. But which parent should I use?

Also note that UCT is not very efficient formula. Using a policy with the AlphaGo formula will be much better.

ok!

flok wrote: ↑

Fri Apr 22, 2022 8:20 pm

Ok. But which parent should I use?

The parent from which you are selecting a move.