RE: [Bug-gnubg] Luck rates ?

[Ian Shaw]

Here are examples of money play and match play from the current
implementation:

                                      yanwolfy             ianshaw

Error rate (total)                    -7.290 ( -8.329)     -3.441 (
-3.549)    
Error rate (per decision)              -34.5 ( -0.039)      -15.9 (
-0.016)    
Equiv. Snowie error rate               -19.4                 -9.2


Luck rate (total)                     +2.929 ( +2.118)     -2.330 (
-4.734)    
Luck rate (per move)                   +15.7 ( +0.011)      -12.3 (
-0.025)   

                                      Bennet               Shaw

Error rate (total)                    -2.808 (-23.283%)    -3.397
(-22.420%)   
Error rate (per decision)              -19.8 ( -0.164%)     -22.6 (
-0.149%)   
Equiv. Snowie error rate                -9.1                -11.0


Luck rate (total)                     -5.198 (-49.754%)    -1.288 (
-8.433%)   
Luck rate (per move)                   -33.3 ( -0.319%)      -8.4 (
-0.055%)   

The current system is confusing because:
1) the bracketed part of the title does not refer to the bracketed part
of the data
2) it is not clear what the first and second (bracketed) values mean
3) "Error rate (total)" and "Luck rate (total)" are not rates at all (as
Christian said). They are simply "Total Errors" and "Total Luck". 
4) the units are not specified, which would be good practice. Totals are
in points for money play or MWC for matches; rates should be in
millipoints per decision for errors or millipoints per move for luck.

I'm undecided on which should be inside the brackets, raw or EMG.
Normalised values are more important for errors, because they give the
student a consistent measurement irrespective of the session length and
match score. This implies that EMG ought to be listed first. 

Raw values are more important for luck, since this is what effects the
outcome - you can't  study to improve your rolling (I hope). This
implies that the raw values should be listed first. 

It would be terribly inconsistent and confusing to vary the display mode
for Errors and Luck, especially since gnubg doesn't state the units of
measurement. Why are raw and moralised values displayed on the same
line? I suspect it's a legacy from the CLI days, when screen space was
limited. If we separate the EMG from the unnormalised values, we get
something like:

                                      yanwolfy             ianshaw

Overall Statistics:
Total Errors (points)                  -7.290                   -3.441

Total Errors as EMG (points)           -8.329                   -3.549

Error rate (mp/decision)               -34.5                  -15.9    
Error rate as EMG (points/decision)    -0.039                   -0.016

Equiv. Snowie error rate               -19.4                -9.2


Luck Statistics:
Total Luck (points)                        +2.929
-2.330
Total Luck as EMG (points)             +2.118                   -4.734
Luck rate (mp/move)                        +15.7                -12.3
Luck rate as EMG (points/move)             +0.011               -0.025

The display occupies four extra lines of text, but they would fit within
the existing area of the gui. It has the advantage of clearly
identifying both measurements.

The units labels are a bit ugly, so maybe we can dispense with them
after all. (Can we do a mouse-over tooltip on this section?) 

It appears that the EMG luck rate and EMG error rate have not been
multiplied by 1000 to convert from points to millipoints per move. There
ought to be consistency here.

I disagree, reluctantly, with MaX's suggestion to use Normalized Equity
(NE). "Normalized equity", is a great description, but "EMG" is the
standard term, so gnubg should adopt it for the sake of the users. Don't
forget that many of them will have more than one bot.

Another recent thread raised by Phoivos Mytilinaios raised the thorny
question of gnubg vs. snowie error rates. I think it's about time we
revisited this issue, and should consider adopting the snowie
implementation, again in the interests of commonality for the bg
community. Even though I have some misgivings about the Snowie way, I
have to acknowledge that it's the lingua franca of discussions about
error rates. No one says, "Neil Kazaross played with a gnubg error rate
of 6.5 in the final". Does showing the gnubg values do anything to
improve the clarity of the information provided?

-- Ian