Pruning in PV nodes

[Henk]

If you apply LMR equally EVERYWHERE, the PV is not given any advantage over any of the other moves.

It certainly does, and the situation is completely analogous to the one you are complaining about. That it may not hurt but may actually be good is another issue.

To quote yourself:

(of course, finding a new PV has to stumble over the reduced search depth problem before it can fail high)

In my words:

LMR makes it harder for other moves to replace the PV move.

Do we really need to discuss the obvious?

But LMR has benefits too. For instance: Less nodes are searched thru. We need percentages, numbers, statistics otherwise you get endless discussions. I think its more like a poker game. The bet is the number of nodes to be searched thru for that branch, the profit a possibly more accurate evaluation.

[bob]

Congratulations, you have just refuted LMR.

How did MY statement refute LMR?

LMR makes it harder for other moves to replace the PV move. I can't make this clearer so I won't try to.

How? If you LMR on the PV as well?

Especially if you do it on the PV, since a move replacing a PV move will have to be a sibling of the PV move.[/quote]

Your point is? Branches along the PV are reduced. Other branches are reduced using identical criteria. There is nothing that favors PV over non-PV in terms of depth, done that way.

Using that reasoning would not a check extension make it harder for any other move to replace the PV move?

No. The check extension will be applied independently of whether a move is the first move at a PV node or a later move.

With LMR you are searching the first move to say depth 20 and a later move will not replace it unless it gives a better score already at depth 19 (or lower). If there are tactics at depth 20, you won't see it. The move that just happens to be good at low depth will tend to stay at the front. (You could counterargue: that first move is more likely to be refuted because you are searching it with more precision. But the same applies when you don't apply reductions to PV nodes that you do apply to non-PV nodes. It is analogous.)
[/quote]

What, EXACTLY, is the "you can't reduce the first move searched" based on? I've not seen such an immutable law. Near the leaves, you know little about move ordering anyway, nothing says you can't reduce those moves that don't meet some sort of criteria for full-depth searching, even if they are the first move searched. I consider reducing to be the opposite of extending. I can extend any move I choose, first searched at a ply or last one searched. Ditto for reductions. But whatever you do, if you do more of it on non-PV nodes, it makes those harder to overtake the actual PV since it is always searched to a deeper depth. I don't see any justification for doing so, to date. Common thinking is "don't reduce first move at a PV node because you don't want to get an n-1 ply search PV when doing an N ply search. I agree. But then the same applies to non-PV nodes. Don't reduce the first one, do reduce the rest, exactly as you did on the PV. Consistency brings stability to the search, which is always a good thing.

If you apply LMR equally EVERYWHERE, the PV is not given any advantage over any of the other moves.

It certainly does, and the situation is completely analogous to the one you are complaining about. That it may not hurt but may actually be good is another issue.

No idea what you are talking about there. If I use the SAME reduction criteria for PV and non-PV it CERTAINLY improves the chances for the non-PV move to sneak up and become best, as opposed to reducing PV moves LESS than the rest, which gives the PV an overall depth advantage it doesn't necessarily deserve.

[bob]

If you apply LMR equally EVERYWHERE, the PV is not given any advantage over any of the other moves.

It certainly does, and the situation is completely analogous to the one you are complaining about. That it may not hurt but may actually be good is another issue.

To quote yourself:

(of course, finding a new PV has to stumble over the reduced search depth problem before it can fail high)

In my words:

LMR makes it harder for other moves to replace the PV move.

Do we really need to discuss the obvious?

We are talking about two DIFFERENT cases.

My case: All nodes are reduced by the same criteria, period.

Your case: PV nodes are reduced LESS than the non-PV nodes. There is a significant difference in those. My original question STILL stands. What is the justification? For example, why not just take the PV nodes as normal and reduce every other move at the root by 2 plies, period. Does THAT not favor the PV even more? That is EXACTLY what treating PV and non-PV nodes differently accomplishes. Maybe (or maybe not) that drastically, or less. But it does favor PV over non-PV if PV is reduced less.

In a perfect tree, searching depth first, left-to-right, ALL the left-most nodes are PV nodes. All of them would be reduced less than any other node in the tree. This reminds me of the old "pv-move extension" which I never saw any benefit to, it always tested as a minor Elo loss when I would try it, yet many proposed it as a good idea. Searching the PV deeper than the rest of the moves is questionable at best.

[syzygy]

Congratulations, you have just refuted LMR.

How did MY statement refute LMR?

LMR makes it harder for other moves to replace the PV move. I can't make this clearer so I won't try to.

How? If you LMR on the PV as well?

Especially if you do it on the PV, since a move replacing a PV move will have to be a sibling of the PV move.

Your point is?

I'm not going to repeat myself forever.

[bob]

Congratulations, you have just refuted LMR.

How did MY statement refute LMR?

LMR makes it harder for other moves to replace the PV move. I can't make this clearer so I won't try to.

How? If you LMR on the PV as well?

Especially if you do it on the PV, since a move replacing a PV move will have to be a sibling of the PV move.

Your point is?

I'm not going to repeat myself forever.

Since you are not answering my question at all, not repeating a useless answer is a good thing...