[computer-go] Is Crazy Stone throwing away half its work?

[Rémi Coulom]

Peter Drake wrote:
> If so, it seems that this would be significant, because (barring the 
> coincidence of choosing C twice in a row) every time a node becomes 
> internal a run is discarded, throwing away nearly half your runs.

Your misunderstanding may come from your confusion of what I call 
internal nodes with nodes stored in memory. The complicated backup and 
selectivity I describe in my paper are applied to internal nodes only. A 
node of the tree is promoted to internal once it has received 81 visits. 
Only a very smal proportion of all the nodes store in memory is promoted 
to internal node.

When the number of simulations that have passed through a node is small, 
the value of the node is the average outcome. So there is no loss of 
information for not storing information about its children. The loss of 
information incurred by my node-allocation scheme starts to have an 
effect only when many random games have passed through this node. I 
expect the loss is much less than 1%.

Rémi