>I have seen this exact behavior when first experimenting with long thinking 
>times in Pachi. When you stop growing the tree, the algorithm degenerates to 
>"delayed" single-level Monte Carlo along the principal variations, with all 
>the MC-without-tree weaknesses.

The pathology definitely depends on the nature of the situation
at the leaves.   For example, one of my 'bots plays hex, and that
robot plays so quickly that the memory limit is reached almost 
immediately.  Yet, continuing to run simulations still improves
the results.  An experiment I just ran at 15 seconds per move, 
hit the memory limit in less than 1 second. If you stop the run after
1 second, the result is a lot weaker than the full 15 second run.

My take-away from this is that small, reasonable-sounding choices
can have huge unintended and unanticipated consequences.  Not really
news to anyone using UCT, but easy to forget.

Computer-go mailing list

Reply via email to