I strongly agree static_shado and Shipwreck: IPR appears to apply to Elo ratings, not Glicko-2 ratings; however, IPR (or some derivative of it) may be a reasonable indicator of what is a "good" or "bad" move.
I think that statement is doubly true for Training (and perhaps Analysis): even without an accurate IPR, I believe the following are possible with a precise IPR:
* Rate the strength of a move relative to other moves on that same ply
* Rate the strength of a move relative to other moves on different plies
* Rate the strength of a puzzle relative to other puzzles of that same category
@static_shadow True, but even without the running average it would be cool. I think its more meaningful than average centipawn loss, and I seam to remember that IPR takes into account the difficulty of the positions (but not sure about it).
In any case, I think it could even attract more players to the site. People are OBSESSED with knowing what their "real" rating would be.
Yeah, agreed with Shipwreck. Here are my thoughts:
1) Doesn't matter if the site uses Glicko. We can give IPR rating (based on Elo) for Tactics Puzzles and analyzed games. I don't think people really care (or are even sensitive to the differences) between Glicko and Elo... they just want some number they can relate to that tells them if they played well or not. I know I would love to input my own OTB games and see what the IPR was for that game. I could see if my IPR changes depending on the time of day, depending on which openings are used, and so on.
It would greatly increase the richness of analysis beyond simply "you made an inaccuracy here."
2) The actual code isn't released, but the math used to produce IPR are in the paper. It should be 'trivially' easy to just write code on the basis of those equations.
3) The point raised about some games not being analyzed, or some games not being analyzed with as much processing depth as others, is a non-issue. It would still give decently accurate IPR scores, using an engine better than the original paper itself, and if that was a problem it would also apply generally to any post-game analysis.
Maybe I'm just confused. If many people are not using the engine, then how would IPR be calculated accurately? I think you have some false idea that all games are run through the engine. They are not. Only games that someone clicks the "generate computer analysis' button go through the engine. The engine is distributed across multiple servers. If the site were to generate this IPR, wouldn't every game have to go through the engine? This would require a SIGNIFICANT increase in servers to distribute the engine analysis. So while the coding itself might be 'trivial' unless I'm completely misunderstanding how this IPR would be calculated, finding the resources to manage it would not be so trivial.
You would be getting an IPR on a PER-GAME BASIS. You click analyze, and as the existing algorithm compartmentalizes your errors into "inaccuracies", "mistakes", and "blunders", you would use that information to assign an IPR based on the average distance between your moves and computer-suggested moves.
Nobody has ever suggested that you need to have every one of your games analyzed for some meta-IPR. You could additionally implement a rolling IPR that is simply the average of all the IPR's you have done generated computer analysis for.
ooooh! OK, so I did completely misunderstand how this IPR thing worked then. Thanks for clearing it up!
If it's just something based on the analysis on a per-game basis (and then potentially a rolling IPR average) then I completely agree that this could actually be a very nice feature.
