Value of IID

[cetormenter]

Recently I have been playing around with IID in Nirvana and have noticed that it isn't as valuable as it once was. Seeing this I decided to collect some statistics and noticed that in almost 95 percent of cases no pv move is even collected from this search! I figured out that if I were to add an extra stipulation that required the hash entry of the position to not have been an alpha bound entry I can get this value to a much more reasonable value of about 6-7 percent. However even with this enhancement it was a complete wash.

I remember when I first implemented IID inside of Nirvana it was worth 10-15 elo but it seems that it's value has declined over time. This is surprising to me as I would have assumed that as my pruning margins became more agresssive move ordering would become more important.

Has anybody else seen similar results?

[jhellis3]

I remember when I first implemented IID inside of Nirvana it was worth 10-15 elo but it seems that it's value has declined over time

How much Elo has Nirvana improved since when you first implemented IID?

There is compression to results the higher one goes (gaining 1 Elo at 3000 is much more difficult than 2000, which is much more difficult than 1000), so if Nirvana has gained a significant amount Elo since then, it is not unreasonable to expect the gain/loss to have been compressed.

[cdani]

For me too, Andscacs has IID for historical reasons. The last time I disabled it was not a loss. But I have not done an extensive test so at some point I will try again.

So is not that the elo gain compresses, but more that probably dissapears in this case.

[jhellis3]

Well you can certainly get compression to 0....

Search improvement (for now) generally means better moves played over a unit of time at the expense of quality per ply. Once moves are of sufficient quality for a given typical time x under testing conditions, then certain optimizations may no longer be relevant. Of course that does not mean they are doing nothing...

An informative test would be fixed ply in comparison to time.

[kbhearn]

Failing to get a PV move out of IID 95% of the time seems very strange.

What are your conditions you check before doing IID?
Is there something about your HT behavior that makes it more likely a fail low node gets overwritten than a fail high?

[kbhearn]

answering myself... i guess if you're doing IID everywhere, then when you decide to IID a fail low node and get no move as part of that process you already IIDed all child fail high nodes and therefore will not need to IID them very often at all and that's why 95% of your IID processes return no move, because the ones that would return a move were often direct descendants of one that wouldn't and thus were already waiting for you in hash

[ymatioun]

I tried implementing IID in Fizbo on several occasions, and never got any improvement.