Balanced Berserker

[edgefigaro]

Curious, what does a balanced berserker start look like? If a char starts at the gate naked, with no items nor skills, but has trog, will they still be broken?

Caveat, I am not proposing a change to berserker. I really like the place in the game berserker has. I'm just curious as to how severe a nerf would balance the background.

[KoboldLord]

Well, let's put it this way.

In the online Crawl Tournament that goes on periodically, there's a specific challenge worth a good deal of points called the 'Nemelex Challenge'. A combination of species and background is randomly generated with no regard to making it viable, and the first person to win with that combo during the tournament gets a bunch of points. Sometimes you get a good combo, but there's some conventional wisdom that can be used to salvage most of the worst possible combinations:

Go Trog. Burn your starting book if you have one.

Mama Trog makes everything better. She knows what you should do.

[Berder]

I think a big part of the advantage of taking trog is he forces you to have a purely physical character. This simplifies things and it means you're doing something you've done many times before instead of trying anything new and uncertain, so that leads to winning.

By the numbers trog doesn't seem to stand out that much; he only has a middle-of-the-pack winrate among non-zealots played by greatplayers who have winrate>10% (restricted to those players in an attempt to weed out speedruns). After the early game, most characters shouldn't be berserking much, and trog's other abilities aren't amazing in comparison to many other gods.

On the other hand, Tedronai's amazing streak was mostly berserkers.

[and into]

... What? What on earth are you talking about, Berder?

I am pretty sure the big advantage to Be is that you start with an ability that gives you (effectively) a bunch of slaying, haste, and +50% HP when you are attacking stuff. That sounds pretty strong to me.

[Bloax]

... What? What on earth are you talking about?

I am pretty sure the big advantage to Be is that you start with an ability that gives you (effectively) a bunch of slaying, haste, and +50% HP when you are attacking stuff. That sounds pretty strong to me.

Berserk is an evokable ring of {+10 slaying, +50% hp, +haste} that leaves you slow and exhausted after the duration expires.
(that's really strong)

[and into]

Evocable abilities sometimes fail, and also take MP, whereas Trog berserk is 100% guaranteed (with no skill investment) and has 0 MP cost, and so it is even better.

[duvessa]

it's not +10 slaying, it's +1d10 to melee damage

[bel]

... What? What on earth are you talking about, Berder?

I am pretty sure the big advantage to Be is that you start with an ability that gives you (effectively) a bunch of slaying, haste, and +50% HP when you are attacking stuff. That sounds pretty strong to me.

Berder was talking about non-zealots.

Anyway, I don't agree with Berder that Trog's other abilities are not that good in comparison to other gods. Brothers in arms kill pretty much everything. And of course the weapon gifts. But of course berserk is the big one. And you can have it on D:1. Pretty much nothing else is needed.

[and into]

Okay, berserk at * is a huge advantage to non-zealots who worship Trog, as well.

[Sar]

Trog's gifts are also actually good.

[gammafunk]

  Code:

<gammafunk> !lg greaterplayers recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu urune<=3 s=god / won o=%
<Sequell> 1061/6204 games for greaterplayers (recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu urune<=3): 115/424x Ashenzari [27.12%], 63/250x Trog [25.20%], 53/221x Ru [23.98%], 76/325x Dithmenos [23.38%], 54/239x Qazlal [22.59%], 76/378x Vehumet [20.11%], 168/871x Okawaru [19.29%], 46/247x Fedhas [18.62%], 86/463x Makhleb [18.57%], 35/189x Nemelex Xobeh [18.52%], 12/65x Yredelemnul [18.46%], 10/62x Elyvilon [16.13%], 71/513x Cheibriados [13.84%], 20/157x Zin [12.74%], 41/329x Kikubaaqudgha [12.46%], 28/236x Sif Muna [11.86%], 12/105x The Shining One [11.43%], 22/225x Gozag [9.78%], 73/905x Xom [8.07%]


No. 2 doesn't look very middle of the pack to me. Filter out boring games, games with non-temple gods, experimental games, and games going for more than 3 runes, and you don't even have to worry about speedruns (and excluding them probably would help Trog's winrate, since he's a favorite speedrun god yet speedruns have higher risk).

[crate]

If Trog's abilities aren't stronger than other gods there is a pretty big problem, since he has two big drawbacks instead of the usual one. (Can't worship a different god, can't cast magic.)

Of course, Trog's abilities are stronger than other gods, so that's not a problem.

I don't personally have problems with Trog's power level. Yes, berserk is very powerful, but it's also powerful in a way that is actually reasonably likely to punish you for making bad plays.

[Berder]

  Code:

<gammafunk> !lg greaterplayers recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu urune<=3 s=god / won o=%
<Sequell> 1061/6204 games for greaterplayers (recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu urune<=3): 115/424x Ashenzari [27.12%], 63/250x Trog [25.20%], 53/221x Ru [23.98%], 76/325x Dithmenos [23.38%], 54/239x Qazlal [22.59%], 76/378x Vehumet [20.11%], 168/871x Okawaru [19.29%], 46/247x Fedhas [18.62%], 86/463x Makhleb [18.57%], 35/189x Nemelex Xobeh [18.52%], 12/65x Yredelemnul [18.46%], 10/62x Elyvilon [16.13%], 71/513x Cheibriados [13.84%], 20/157x Zin [12.74%], 41/329x Kikubaaqudgha [12.46%], 28/236x Sif Muna [11.86%], 12/105x The Shining One [11.43%], 22/225x Gozag [9.78%], 73/905x Xom [8.07%]


No. 2 doesn't look very middle of the pack to me. Filter out boring games, games with non-temple gods, experimental games, and games going for more than 3 runes, and you don't even have to worry about speedruns (and excluding them probably would help Trog's winrate, since he's a favorite speedrun god yet speedruns have higher risk).

That's a flawed query. The practice of restricting to urune<=3 doesn't do what you think it does. Suppose you take a god pre-lair, such as Vehumet, and plan to go for 15 runes or bust. If you die with <=3 runes, your death counts against Vehumet in your query. But if you win, it never counts for Vehumet in that query, because if you win you'll have had 15 runes! Your query therefore has a bias for gods that people 3-rune with, and a bias against gods that people take pre-lair but continue to do more runes with. This bias favors Trog.

"But," you might protest, "what about the gods that people switch to for the extended endgame?" Yes, that is a problem, but there are other ways to compensate for it. The simplest way is just to discount the traditional extended endgame gods (makhleb, tso, zin, jiyva). A more complex query could try to look at when you worshiped the god and whether you switched. It would also be more accurate to look at / urune>=3 rather than / won, so that deaths in extended don't count against a god.


If we redo your query without that urune<=3 this is what we see:

  Code:

<halberd> !lg greaterplayers recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu s=god / won o=%
<Sequell> 1653/6890 games for greaterplayers (recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu): 74/182x The Shining One [40.66%], 165/484x Ashenzari [34.09%], 146/457x Vehumet [31.95%], 77/248x Ru [31.05%], 113/364x Dithmenos [31.04%], 81/268x Qazlal [30.22%], 82/275x Trog [29.82%], 150/539x Makhleb [27.83%], 44/182x Zin [24.18%], 48/204x Nemelex Xobeh [23.53%], 216/923x Okawaru
<Sequell> [23.40%], 16/70x Yredelemnul [22.86%], 58/261x Fedhas [22.22%], 74/364x Kikubaaqudgha [20.33%], 13/65x Elyvilon [20.00%], 112/565x Cheibriados [19.82%], 47/259x Sif Muna [18.15%], 33/239x Gozag [13.81%], 104/941x Xom [11.05%]

We see that Ash, Veh, Ru, Dith, and Qaz all come out ahead of Trog. None of those gods are often switched to for extended; the reason they came behind in your query is that flawed urune<=3 condition.


Also, you're right that speedrunners are a problem and may make Trog place unfairly low. Let's redo that query with "tenpercenters," who are greatplayers with winrate>=10%. tenpercenters probably don't do very much speedrunning - at any rate less speedrunning than greatplayers who aren't tenpercenters.

  Code:

<halberd> !lg tenpercenters recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu s=god / won o=%
<Sequell> 871/2155 games for tenpercenters (recent !boring !experimental cls!=be|ak|he|dk god!= god!=jiyva|beogh|lugonu): 48/76x The Shining One [63.16%], 34/67x Ru [50.75%], 67/133x Dithmenos [50.38%], 78/155x Ashenzari [50.32%], 80/170x Vehumet [47.06%], 32/69x Trog [46.38%], 90/199x Makhleb [45.23%], 35/79x Fedhas [44.30%], 23/52x Zin [44.23%], 6/14x Elyvilon [42.86%], 31/74x Qazlal [41.89%], 24/59x
<Sequell> Nemelex Xobeh [40.68%], 14/35x Gozag [40.00%], 52/132x Kikubaaqudgha [39.39%], 111/284x Okawaru [39.08%], 11/32x Yredelemnul [34.38%], 40/123x Cheibriados [32.52%], 30/103x Sif Muna [29.13%], 65/299x Xom [21.74%]

Trog still places behind Ru, Dith, Ash, and Veh. With the / urune>=3 condition instead of / won, he would also place behind Mak (though we should possibly discount mak, as he may count as a god many people switch to).



On the subject of Trog's abilities other than early game berserk, I don't think they are hugely strong compared to other gods. Brothers in arms are nice, sure. Okawaru more than doubles your combat effectiveness when his abilities are used. Dithmenos helps you completely avoid fights, and gives you passive boosts (umbra and shadow mimic) that might amount to as much as +50% combat effectiveness all the time on some characters. A mage with veh is able to stay in a fight much longer than a mage without veh. Not only does mak give you silly amounts of healing, he can summon greater demons that are just as good if not better than BiA (unless hostile). As far as "kill everything" goes, Qaz's disaster area is far more effective in the right scenario than a BiA or two. And then there's Ru.

[Sar]

well I guess it's settled then, Qaz is better than Trog

(lol)

[and into]

I hesitate to ask, but... What do you think those numbers prove?

[crate]

You really want to use some sort of lm query here instead of lg, probably, since then you're not excluding Trog->othergod (Berder's query) and simultaneously not including othergod->thirdgod (gamma's query).

For instance, perhaps you compare

  Code:

<crate> !lm * recent playable !boring rune !ak !be !ck !dk !--he !experimental s=god
<Sequell> 82733 milestones for * (recent playable !boring rune !ak !be !ck !dk !--he !experimental): 13731x Okawaru, 10940x Vehumet, 8833x The Shining One, 6919x Cheibriados, 6872x Ashenzari, 6224x Makhleb, 4098x Qazlal, 3663x Dithmenos, 3217x Sif Muna, 2766x Zin, 2158x Kikubaaqudgha, 2101x Ru, 1922x, 1831x Trog, 1496x Beogh, 1356x Gozag, 1216x Nemelex Xobeh, 1102x Jiyva, 871x Fedhas, 519x Yredelemnul,
<Sequell> 383x Elyvilon, 264x Xom, 251x Lugonu

or alternatively

  Code:

<crate> !lm * recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3 s=god
<Sequell> 49442 milestones for * (recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3): 11358x Okawaru, 6186x Vehumet, 3895x Ashenzari, 3800x Cheibriados, 3340x Makhleb, 3068x Qazlal, 2574x Dithmenos, 1978x Sif Muna, 1728x The Shining One, 1467x Trog, 1370x Kikubaaqudgha, 1319x Ru, 1249x, 1134x Beogh, 972x Gozag, 913x Nemelex Xobeh, 704x Zin, 692x Fedhas, 642x Jiyva, 460x Yredelemnul,
<Sequell> 292x Elyvilon, 155x Xom, 146x Lugonu

to

  Code:

<crate> !lg * recent playable !boring !ak !be !ck !dk !--he !experimental s=god
<Sequell> 681359 games for * (recent playable !boring !ak !be !ck !dk !--he !experimental): 483196x, 41765x Okawaru, 24905x Vehumet, 14862x Cheibriados, 14132x Makhleb, 14130x Sif Muna, 12783x Ashenzari, 12220x Dithmenos, 11922x Qazlal, 7759x The Shining One, 7145x Kikubaaqudgha, 6337x Nemelex Xobeh, 5568x Trog, 4652x Ru, 4064x Gozag, 3838x Beogh, 3064x Fedhas, 2875x Zin, 1844x Yredelemnul, 1623x Xom,
<Sequell> 1309x Elyvilon, 920x Jiyva, 446x Lugonu

and see what that gives you.
For instance, Trog is responsible for about 2% of all runes in <1% of all games; or Trog is responsible for about 3% of all first-three-runes in <1% of all games. I'm too lazy to go further or get any conclusions (most deaths of course are godless, so the fact that Trog has a larger share of runes than of games is to be expected). I will suggest that this sort of query is probably superior to what either Berder or gammafunk are doing above, though. Probably you can improve my idea (and of course I did no filtering for player quality at all); perhaps you want to compare to a god.worship milestone count instead of to a total number of games count. But looking at god at rune milestone time is, I will suggest, more representative than looking at end-game god for what you are trying to do, and that's the real key.

(h/t to ebarrett for checking rune milestones as an idea)

---

That said I think it's pretty clear that Trog is a strong god.

[Berder]

Crate, you have a good idea.

  Code:

<crate> !lm * recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3 s=god
<Sequell> 49442 milestones for * (recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3): 11358x Okawaru, 6186x Vehumet, 3895x Ashenzari, 3800x Cheibriados, 3340x Makhleb, 3068x Qazlal, 2574x Dithmenos, 1978x Sif Muna, 1728x The Shining One, 1467x Trog, 1370x Kikubaaqudgha, 1319x Ru, 1249x, 1134x Beogh, 972x Gozag, 913x Nemelex Xobeh, 704x Zin, 692x Fedhas, 642x Jiyva, 460x Yredelemnul,
<Sequell> 292x Elyvilon, 155x Xom, 146x Lugonu


I like this query. We don't want to evaluate extended performance so urune<=3 is good here. It doesn't have the urune<=3 problem that gammafunk's game query had.

  Code:

!lg * recent playable !boring !ak !be !ck !dk !--he !experimental s=god

This query is more problematic since it's only counting the god you have at the end of the game, and people can switch. I think it would be better to compare to just the low level god joining milestones:

  Code:

<halberd>!lm * recent playable !boring !zealot !experimental god.worship xl<=12 s=noun
<Sequell> 200653 milestones for * (recent playable !boring !zealot !experimental god.worship xl<=12): 43114x Okawaru, 24988x Vehumet, 14956x Cheibriados, 14208x Sif Muna, 14089x Makhleb, 12894x Ashenzari, 12403x Dithmenos, 12057x Qazlal, 7311x the Shining One, 7230x Kikubaaqudgha, 6438x Nemelex Xobeh, 5748x Trog, 4834x Ru, 4255x Gozag, 3895x Beogh, 3172x Fedhas, 2813x Zin, 1969x Yredelemnul, 1798x Xom,
<Sequell> 1551x Elyvilon, 568x Jiyva, 362x Lugonu

Now let's calculate one divided by the other (I did something similar to this for DrKe's god rating thread, so I still have the code):

  Code:

Xom 0.0862068965517 ( 155 / 1798 )
Sif Muna 0.139217342342 ( 1978 / 14208 )
Nemele Xobeh 0.141814228021 ( 913 / 6438 )
Elyvilon 0.188265635074 ( 292 / 1551 )
Kikubaaqudgha 0.18948824343 ( 1370 / 7230 )
Dithmenos 0.207530436185 ( 2574 / 12403 )
Fedhas 0.21815889029 ( 692 / 3172 )
Gozag 0.228437132785 ( 972 / 4255 )
Yredelemnul 0.233621127476 ( 460 / 1969 )
the Shining One 0.236356175626 ( 1728 / 7311 )
Makhleb 0.237064376464 ( 3340 / 14089 )
Vehumet 0.247558828238 ( 6186 / 24988 )
Zin 0.250266619268 ( 704 / 2813 )
Cheibriados 0.25407863065 ( 3800 / 14956 )
Qazlal 0.254457991208 ( 3068 / 12057 )
Trog 0.255219206681 ( 1467 / 5748 )
Okawaru 0.263441109616 ( 11358 / 43114 )
Ru 0.272858916012 ( 1319 / 4834 )
Beogh 0.291142490372 ( 1134 / 3895 )
Ashenzari 0.302078486118 ( 3895 / 12894 )
Lugonu 0.403314917127 ( 146 / 362 )
Jiyva 1.13028169014 ( 642 / 568 )


So we see, trog is better than average but not quite at the head. Oka, Ru, and Ash score higher. Also, Zin, Chei, and Qaz are very close behind Trog and could be considered equally good by this measure.

Also, a big flaw with the above queries is that they are across all players. The average player doesn't know how to play and trog is simple, so they have an easier time with him. Let's repeat the experiment with just tenpercenters:

  Code:

<halberd> !lm tenpercenters recent playable !boring !zealot !experimental god.worship xl<=12 s=noun
<Sequell> 1917 milestones for tenpercenters (recent playable !boring !zealot !experimental god.worship xl<=12): 320x Okawaru, 183x Makhleb, 175x Vehumet, 150x Ashenzari, 134x Kikubaaqudgha, 134x Dithmenos, 122x Cheibriados, 103x Sif Muna, 84x Fedhas, 76x Qazlal, 73x Trog, 66x Ru, 58x Nemelex Xobeh, 45x Zin, 39x the Shining One, 38x Gozag, 36x Yredelemnul, 22x Beogh, 21x Elyvilon, 16x Xom, 13x Lugonu, 9x
<Sequell> Jiyva
<halberd> !lm tenpercenters recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3 s=god
<Sequell> 3061 milestones for tenpercenters (recent playable !boring rune !ak !be !ck !dk !--he !experimental urune<=3): 488x Okawaru, 301x Vehumet, 274x Makhleb, 252x Ashenzari, 215x Dithmenos, 201x Kikubaaqudgha, 156x Cheibriados, 144x, 140x Fedhas, 120x Trog, 117x Qazlal, 115x Sif Muna, 108x Ru, 82x Nemelex Xobeh, 59x The Shining One, 56x Zin, 48x Yredelemnul, 43x Beogh, 42x Jiyva, 41x Gozag, 24x
<Sequell> Lugonu, 19x Elyvilon, 16x Xom

with these results:

  Code:

Elyvilon 0.904761904762 ( 19 / 21 )
Xom 1.0 ( 16 / 16 )
Gozag 1.07894736842 ( 41 / 38 )
Sif Muna 1.11650485437 ( 115 / 103 )
Zin 1.24444444444 ( 56 / 45 )
Cheibriados 1.27868852459 ( 156 / 122 )
Yredelemnul 1.33333333333 ( 48 / 36 )
Nemele Xobeh 1.41379310345 ( 82 / 58 )
Makhleb 1.49726775956 ( 274 / 183 )
Kikubaaqudgha 1.5 ( 201 / 134 )
the Shining One 1.51282051282 ( 59 / 39 )
Okawaru 1.525 ( 488 / 320 )
Qazlal 1.53947368421 ( 117 / 76 )
Dithmenos 1.60447761194 ( 215 / 134 )
Ru 1.63636363636 ( 108 / 66 )
Trog 1.64383561644 ( 120 / 73 )
Fedhas 1.66666666667 ( 140 / 84 )
Ashenzari 1.68 ( 252 / 150 )
Vehumet 1.72 ( 301 / 175 )
Lugonu 1.84615384615 ( 24 / 13 )
Beogh 1.95454545455 ( 43 / 22 )
Jiyva 4.66666666667 ( 42 / 9 )


Trog still does fairly well, though not at the head. Trog is behind fedhas, ash, and veh, and on a par with ru and dith.

[adozu]

"So we see, trog is better than average but not quite at the head. Oka, Ru, and Ash score higher. Also, Zin, Chei, and Qaz are very close behind Trog and could be considered equally good by this measure."

"So we see, trog is better than average ... Also, Chei ... could be considered equally good by this measure."

wait, what?

i like chei, but i think any measure that indicates chei is about as good as trog is flawed somehow.

[Berder]

... among "average" players who barely know when they need to run away anyway, chei's walking speed apparently isn't as large a disadvantage.

The later result among tenpercenters is more relevant, and chei ranks lower there.

[n1000]

So Jiyva players get 4.6 runes while picking up their three first runes per game

An issue is that while these issues do have some overlap, of course, they're not tracking the same games. Another exception, stemming from anecdote, I'll admit, is that a tenpercenter might bias his choice of god based on strength of the character. It is frequently claimed that book starts are stronger than melee starts, and I avoid Trog on these characters. Furthermore, I am a tenpercenter that speedruns occasionally and Trog is on the short list of gods I am willing to worship in these games. I also tend to pick Trog on weaker characters because I believe that god to be relatively more powerful than the others in many of these sorts of situations.

There is a canonical example from epidemiology in which a statistician attempts to debunk some conventional wisdom by pointing out that, despite the dry air, Arizona has a much higher TB rate than Washington. Maybe humid atmospheres actually hinder virus survival! Of course the real reason, according to the parable*, is that Arizona has a higher TB rate specifically because dry air helps people infected with tuberculosis! TB patients move to the driest states as part of their treatment, and thus skew the statistics into something misleading.

*

Spoiler: show

The actual cause of the relatively high incidence of TB in Arizona are more complicated. It's related to the generally higher incidence in Hispanic populations, that a large fraction of US cases originate in Latin America, and that Arizona is home to so many immigrants from these countries, among other factors

[Berder]

So Jiyva players get 4.6 runes while picking up their three first runes per game

I believe this anomaly is due to people taking Jiyva as their first god in lair at xl>=13. (or maybe due to people switching fedhas->jiyva before they have three runes). It shouldn't be an issue for the rest of the results.

An issue is that while these issues do have some overlap, of course, they're not tracking the same games. Another exception, stemming from anecdote, I'll admit, is that a tenpercenter might bias his choice of god based on strength of the character. It is frequently claimed that book starts are stronger than melee starts, and I avoid Trog on these characters. Furthermore, I am a tenpercenter that speedruns occasionally and Trog is on the short list of gods I am willing to worship in these games. I also tend to pick Trog on weaker characters because I believe that god to be relatively more powerful than the others in many of these sorts of situations.

Both of those (weaker chars taking trog, and trog speedrunning) may still be factors. Fedhas is the more typical god to use for weak chars though, and Fedhas ranks above Trog in the results.

It is in theory possible to control for this stuff too but I think I would need to download all the milestones and games and analyze them with my own tools. Sequell queries seem inadequate.

I did another query among tenpercenters that looks at only milestones for the 3rd rune gotten ignoring the 1st and 2nd runes. I think this is more relevant (after all what does it matter if you get 2 runes and then die?) These were the results.

  Code:

Elyvilon 0.285714285714 ( 6 / 21 )
Sif Muna 0.31067961165 ( 32 / 103 )
Xom 0.3125 ( 5 / 16 )
Gozag 0.342105263158 ( 13 / 38 )
Cheibriados 0.385245901639 ( 47 / 122 )
Zin 0.4 ( 18 / 45 )
Yredelemnul 0.416666666667 ( 15 / 36 )
Nemele Xobeh 0.431034482759 ( 25 / 58 )
Kikubaaqudgha 0.432835820896 ( 58 / 134 )
Fedhas 0.440476190476 ( 37 / 84 )
Qazlal 0.460526315789 ( 35 / 76 )
Okawaru 0.478125 ( 153 / 320 )
Trog 0.479452054795 ( 35 / 73 )
Makhleb 0.48087431694 ( 88 / 183 )
Ru 0.5 ( 33 / 66 )
Dithmenos 0.507462686567 ( 68 / 134 )
Vehumet 0.537142857143 ( 94 / 175 )
Ashenzari 0.553333333333 ( 83 / 150 )
the Shining One 0.564102564103 ( 22 / 39 )
Beogh 0.636363636364 ( 14 / 22 )
Lugonu 0.769230769231 ( 10 / 13 )
Jiyva 2.22222222222 ( 20 / 9 )


Trog and fedhas fall down a bit more, here, perhaps partly because they aren't as strong towards midgame and endgame, and partly because people are already switching away from them by that time. I'm guessing TSO's high placing is due to people switching to TSO already before the third rune.

[Pereza0]

Trog is indeed strong, yes.

But I think he also has the advantage of being more straightforward than any god, this is accomplished through restrictions but in a game with the amount of diversity Crawl has this is not too bad, helps you avoid getting sidetracked. His powers are really good, but other gods also have comparably great (though maybe less obvious) early powers.

With Magic and Invocations out of the equation your only worry is basically your melee prowess and maybe some Evocations or ranged. It is hard to mess that up, and I think this is part of the reason Trog works so well, especially for newbies. He is more than a handful of abilities, by worshiping him you are basically declaring what a big chunk (or all) of your game is going to look like which is not really the case with most gods.

Nerfing Berserker by removing skills or starting weaponry would only serve to annoy and to trap newbies, I would rather see the background removed than having something like that.

Bersekers are very powerful but very focused. I think that is interesting, and I would say having some things be more powerful than other things is fine as long as it results in interesting stuff.

[Pereza0]

i like chei, but i think any measure that indicates chei is about as good as trog is flawed somehow.

Yep.

I think it is hard to get any idea by looking at player statistics.

Any new player will give Trog a shot and die while learning.

A guy who picks Chei is more likely to be more experienced.

[Berder]

Trog is indeed strong, yes.

But I think he also has the advantage of being more straightforward than any god, this is accomplished through restrictions but in a game with the amount of diversity Crawl has this is not too bad, helps you avoid getting sidetracked. His powers are really good, but other gods also have comparably great (though maybe less obvious) early powers.

With Magic and Invocations out of the equation your only worry is basically your melee prowess and maybe some Evocations or ranged. It is hard to mess that up, and I think this is part of the reason Trog works so well, especially for newbies. He is more than a handful of abilities, by worshiping him you are basically declaring what a big chunk (or all) of your game is going to look like which is not really the case with most gods.

Nerfing Berserker by removing skills or starting weaponry would only serve to annoy and to trap newbies, I would rather see the background removed than having something like that.

Bersekers are very powerful but very focused. I think that is interesting, and I would say having some things be more powerful than other things is fine as long as it results in interesting stuff.

I agree entirely. A big part of why Trog is strong is that he forces simple and focused characters.

[Sar]

Didn't you just prove Trog is pretty mediocre? I mean Ash is clearly better and it's not like Ash is a particularly good god (he doesn't do anything until Lair on most characters).

[Berder]

I think the results show that Ash is actually very strong. He doesn't do anything early on but then he helps a lot, with the skill boosts and enemy detection. Also, the kind of character who takes Ash plays a role in Ash's standing: generally hybrids or mages take Ash. So it speaks to the strength of being a hybrid rather than only melee.

Also, bear in mind there were two factors (possible tenpercenters who are speedrunning with trog despite being tenpercenters - and weak characters taking trog to bolster their early game survival) - which could increase Trog's standing a bit if they were controlled for. And Trog didn't rank poorly, he just didn't rank at the very top. "Trog is not overpowered" is the conclusion I would draw from the results we've examined.

[Lasty]

!lg and !lm are very interesting, but they have nowhere close the level of scientific rigor Berder is attempting to vest in them. Even if you filter out the obvious issues, and some progress was made on that front over the course of this thread, you still can't control for possibilities like "greaterplayers have a tendency to worship Trog when they're feeling lazy", or "greaterplayers have a tendency to worship Trog when their race/background is weak" or "greaterplayers tend to get bored while playing Trog and make more mistakes than usual", or countless other confounding scenarios. These aren't purely hypothetical: 2 and 3 apply to me, at the very least.

Luckily, querying a data set full of confounding variables isn't our only way to evaluate the strength of gods. We can draw on personal experience Trog offers strong abilities at low piety levels and (almost) guarantees access to a super-powered weapon early in the game. It's trivially true that Trog is very, very strong, and that starting with a very, very strong god is very, very good. Is Trog the only very, very strong god? No, and I don't think anyone is claiming so. Trog does give you a complete package to ensure you win, though: Very powerful ability early in the game? Check. A power that can take down dangerous things without personal danger? Check. A strong, cheap defensive power? Check. Provides a game plan that takes you from the start of the game to the end without trouble? Check. The only thing Trog doesn't offer (arguably) is an incredibly powerful "cannot die no matter how badly I screwed up" ability like Sanctuary.

Is Trog the strongest god? In order to determine the answer to that, we'd need to set up reasonable and agreed-upon criteria for what makes a god strong. I don't know if Trog is the absolute strongest, but Trog is definitely a contender.

[Berder]

For all their flaws, the milestone and game data are the only way to reach conclusions that have a semblance of objectivity. Otherwise you just have somebody saying one thing, and another person saying another thing, and no way to reconcile the two views or see to what extent one person is right.

It's always better to look at whatever evidence there is, even if it's not quite the evidence you want.

[basil]

Except when the "evidence" is so sketchy as to be useless.

[Berder]

Yes, except if the data is truly useless. The lg and lm data is useful, though susceptible to various biases which an experienced player can guess at and try to correct for. A good interpretation of the data can at least get you in the ballpark of the truth. We can say that Trog is not head and shoulders above every other god. We can say there are many gods that seem like they might be as good or better than Trog, as far as winrate is concerned.

Actually, it looks like Trog is not used much for speedruns on a non-berserker. Our previous efforts to correct for that weren't necessary. Take a look.

  Code:

<halberd> !lg * won urune=3 turn<40000 !be s=god
<Sequell> 108 games for * (won urune=3 turn<40000 !be): 20x Elyvilon, 18x Jiyva, 11x Okawaru, 10x Ashenzari, 9x Makhleb, 7x Cheibriados, 5x Dithmenos, 4x Vehumet, 4x Zin, 4x Sif Muna, 3x Lugonu, 3x Trog, 3x, 2x Ru, Qazlal, Yredelemnul, Xom, Fedhas, Beogh

Only three non-berserkers who won in under 40k turns with 3 runes chose Trog.
Or we could look at 3 rune wins <50k turns, with similar results.

  Code:

<halberd> !lg * won urune=3 turn<50000 !be s=god
<Sequell> 252 games for * (won urune=3 turn<50000 !be): 34x Ashenzari, 31x Jiyva, 28x Okawaru, 27x Makhleb, 25x Elyvilon, 18x Cheibriados, 12x Trog, 10x Lugonu, 10x Dithmenos, 8x Yredelemnul, 7x Ru, 6x Vehumet, 6x Sif Muna, 4x Fedhas, 4x, 4x Zin, 4x Qazlal, 3x The Shining One, 3x Beogh, 3x Xom, 2x Nemelex Xobeh, 2x Kikubaaqudgha, Wulndraste

[Lasty]

For all their flaws, the milestone and game data are the only way to reach conclusions that have a semblance of objectivity.

The semblance of objectivity without any actual objectivity is far worse than simply admitting that objective correctness is not attainable. In this case, objectivity is not attainable.

[Siegurt]

Well, and unfortunately, the type of evidence we have is so varied that it can't be viewed in a way that objectively lets you see the conclusions you're trying to draw directly from the data itself. The problem with statistics is that even if the data is objective, the observer is not, and whenever you try to derive conclusions from data, you're subject to both observer bias and (in this case) query bias.

Another problem with this *particular* set of conclusions, is that it's highly subjective, that is to say, Trog's powers may be more useful for some players than for others, depending on their playstyle. Yet *another* problem is that different gods are differently-powerful for different areas of the game, resulting in a non-smooth, dissimilar curve for power level of gods for different parts of the game.

I submit that in this case using queries to attempt to prove Trog's power level compared to other gods is no more or less reliable than collecting user anecdotes, expressly because "god's power level" isn't something that's representable as an abstract or is derivable directly from any of our given data. It's not possible to separate a given character's success as it is attributable to the god vs not. After all, one player could worship a god, do exceptionally well, and never use any of his powers or gifts, and another one could rely heavily on a god and die early, in circumstances not related to the influence of the god at all.

[Berder]

nope, data > anecdotes.

The data we have here is actually pretty good.

[Sandman25]

I suspect some greaterplayers can choose Trog when they want to have fun instead of winning. I mean the god allows to recklessly tab things which can result in stupid deaths.

[Berder]

I suspect some greaterplayers can choose Trog when they want to have fun instead of winning. I mean the god allows to recklessly tab things which can result in stupid deaths.

One way to correct for that would be to look only at games that could potentially be part of a streak - assuming players want streaks.

In my experience it's easier to recklessly tab things when I'm with mak or chei than trog. Trog gives you a good weapon, but you don't want to be berserking the whole game and he doesn't give you passive buffs.

[n1000]

For all their flaws, the milestone and game data are the only way to reach conclusions that have a semblance of objectivity. Otherwise you just have somebody saying one thing, and another person saying another thing, and no way to reconcile the two views or see to what extent one person is right.**

This isn't actually the case (obviously) but I can't actually find the survey that backs up my opinion-based claim on the utility of arguments. Let me gather some quick data: thank this post if you agree.

But seriously, there is a very basic flaw with your reasoning in the post above, and there are two other examples of claims in other posts which I see as being particularly poorly justified. I noticed these instances thanks to my immense capacity for observation and deduction, so instead of pointing them out, I'll leave their discovery as an exercise in reasoning for the reader*

However, there are some excellent points you bring up as well. Don't take this post to be a personal attack (except in the sense that every post I have ever made is a personal attack against all of my readers--but I don't expect more than one person to understand this point as it requires some familiarity with philosophy) but precisely as me taking advantage of this discussion to draw my readers' attention towards fallacies whose discovery might be fun or educational.

* Some hints: 1: DCSS is a video game, and an open source one, at that. 2: What makes Trog strong? 3: Objectivity of data versus the conclusions herein drawn from the data.

** Bonus question: Assuming good faith and proper reasoning among the interlocutors,*** why is

you just have somebody saying one thing, and another person saying another thing, and no way to reconcile the two views or see to what extent one person is right.

false? It is false in the general case, not just in the context of our discussion.

Please turn your answers in by midnight on the date listed on the syllabus.

*** which we don't have

anyway, bye all

[Berder]

Do you read hpmor, n1000?

[n1000]

I read a chapter once, on recommendation from another crawler, incidentally

[Berder]

Oh, something about what you said reminded me of it. Anyway, I think it's pointless to argue over how to define "best." The data says what it says: among tenpercenters (i.e. fairly competent players who don't speedrun much), trog's rate of getting 3 runes is on a par with several other gods, and not clearly ahead of them. Each person can define "best" as they please, informed by that data.

I'm not really sure what you're getting at in your long post. In practice, opinion divorced from data doesn't convince people, and nor should it; opinion is not evidence.

[Pereza0]

Berserkers

[and into]

Correctly gathered data is always an accurate measure of something. But what it is a measure of is not always clear. And if you don't actually understand the object of inquiry then data (by itself) is just as likely to mislead you as anything else. (Both science and pseudoscience often rely on technically accurate statistics. The difference is that science uses such data honestly, while pseudoscience uses it dishonestly. There often is not any way to determine one from the other a priori; you have to actually look at the specific claims and apply critical thinking.) The point is, data requires interpretation. This is why statisticians and scientists have a job, and not just the programmers who write the algorithms for various math programs.

Several reasonable interpretations different from your own have been offered, Berder, which you have (wrongly) dismissed as purely anecdotal. (An interpretation of data and analysis of confounding factors is not just an "anecdote." There are different forms of (legitimate, reason-based) evidence and argumentation than "data" and "anecdote.")

This is CYC, so whatever, but when you post another block of inquiries or fsim or whatever in DCA or GDD, please drop the lame "I post teh data, so I am right" shtick. Posting the data and presenting an argument is, of course, fine.



Anyway, I'm willing to accept that I might overrate Trog, I guess (obviously Trog is extremely good/strong, but it is still possible to overrate something good/strong). Before I decide on that, though, I would want to see a different kind of data. (I don't know how to do the request statistics wizardry you all seem so adept at.) Basically, with the usual exceptions (tenpercenter, etc.), I'd want to see (if possible) %win for three runes with species that straightforwardly—not a ton of gimmicks—make sense for Trog (I'd say Mi, Hu, HO, just to start, obviously there are others) of backgrounds that do not start with books or a god, and who worship Trog. Then I would like to see the win percentage for the same set, except worshiping other gods (I would suggest analyzing each one individually, with the exception of Sif, Vehumet, and Kiku, as those gods are more strongly slanted toward book starts, whereas Mi and HO as species are not, so it doesn't seem useful for this particular comparison).

That would be a lot more work, and I don't know if that's even possible, but that would give a much better and more direct comparison for how strong Trog is on characters that do not start with a book or god and which are not challenge characters.

You see, the thing you keep ignoring (and it is a big thing, not a small thing) is that when people play games of Crawl online—even very good players—they are making choices based on a number of factors. The fact that Be is a start complicates matters greatly; when you want to play Trog, you just start a Be the vast majority of the time. Not some of the time—most of the time. I think it follows that a greater percentage (not all, but proportionately, a greater percentage) of characters that do not start as Be, but then worship Trog, are examples of games where some weird conditions are in play, when compared to gods that you cannot start with. That's not the only confounding factor, but I expect it could be one major factor.

The things you keep brushing off as "minor" are actually things that would be grounds for not having strong confidence in the findings. That is, assuming one is approaching the data in a genuinely scientific spirit.

[Greyr]

Sorry, I tried reading this thread but zoned out after

[code][/code]

[all before]

data > anecdotes

Until you can statistically prove statistics are better than other forms of argument without making any (non-statistical) recourse to those forms of argument, I refuse to accept your claim here as valid.

[Berder]

Correctly gathered data is always an accurate measure of something. But what it is a measure of is not always clear. And if you don't actually understand the object of inquiry then data (by itself) is just as likely to mislead you as anything else. (Both science and pseudoscience often rely on technically accurate statistics. The difference is that science uses such data honestly, while pseudoscience uses it dishonestly. There often is not any way to determine one from the other a priori; you have to actually look at the specific claims and apply critical thinking.) The point is, data requires interpretation. This is why statisticians and scientists have a job, and not just the programmers who write the algorithms for various math programs.

I think I understand the object of inquiry fairly well, being a greateroctopode. Nor could you accuse me of applying some prepackaged algorithm without thought of what it means to the game. Much of this thread has involved me and others trying different ways to analyze the data that could correct for various confounding factors.

Several reasonable interpretations different from your own have been offered, Berder, which you have (wrongly) dismissed as purely anecdotal. (An interpretation of data and analysis of confounding factors is not just an "anecdote." There are different forms of (legitimate, reason-based) evidence and argumentation than "data" and "anecdote.")

The interpretations of the data are not anecdotes and I didn't mean that. Anecdotes would be things like "I've played trog a few times and he seemed very strong to me." Also, things like "Look at how powerful berserk and BiA is! They can easily take down X Y Z" might not exactly be anecdotes, but they aren't data-supported either, because they don't take into account the strengths of other gods.

Anyway, you can compare and contrast different god abilities to try to argue how one is stronger than the other, but gods are not directly comparable that way. It's not like one god lets you summon giants and invoke firestorm and another one lets you summon ogres and invoke fireball. They all have their strengths and weaknesses in different situations. There's no better way to compare them overall than looking at the statistics.

Anyway, I'm willing to accept that I might overrate Trog, I guess (obviously Trog is extremely good/strong, but it is still possible to overrate something good/strong). Before I decide on that, though, I would want to see a different kind of data. (I don't know how to do the request statistics wizardry you all seem so adept at.) Basically, with the usual exceptions (tenpercenter, etc.), I'd want to see (if possible) %win for three runes with species that straightforwardly—not a ton of gimmicks—make sense for Trog (I'd say Mi, Hu, HO, just to start, obviously there are others) of backgrounds that do not start with books or a god, and who worship Trog. Then I would like to see the win percentage for the same set, except worshiping other gods (I would suggest analyzing each one individually, with the exception of Sif, Vehumet, and Kiku, as those gods are more strongly slanted toward book starts, whereas Mi and HO as species are not, so it doesn't seem useful for this particular comparison).

That would be a lot more work, and I don't know if that's even possible, but that would give a much better and more direct comparison for how strong Trog is on characters that do not start with a book or god and which are not challenge characters.

Well, you start running into problems of sample size if you divide it up that finely. For instance, there are only 89 games for tenpercenters who are Mi or Hu-- or HO, and Fi or As or Gl, and only 3 of those chose trog. However I can link to this: viewtopic.php?f=17&t=14701 showing the easiest and hardest classes to win according to an Elo analysis. OpTm has the highest elo (hardest), GrFi has the lowest elo. Some mages are also very easy, but mostly tanky mages like GrEE or DDNe. I would say that a strong physical race paired with Trog is quite powerful. But a strong physical race paired with almost any god is powerful.

You see, the thing you keep ignoring (and it is a big thing, not a small thing) is that when people play games of Crawl online—even very good players—they are making choices based on a number of factors. The fact that Be is a start complicates matters greatly; when you want to play Trog, you just start a Be the vast majority of the time. Not some of the time—most of the time. I think it follows that a greater percentage (not all, but proportionately, a greater percentage) of characters that do not start as Be, but then worship Trog, are examples of games where some weird conditions are in play, when compared to gods that you cannot start with. That's not the only confounding factor, but I expect it could be one major factor.

Well, you can look at what characters took trog.

  Code:

<halberd> !lg tenpercenters trog !zealot s=char
<Sequell> 380 games for tenpercenters (trog !zealot): 23x OgWn, 10x OgHu, 8x MiFi, 7x DsAr, 6x TeGl, 6x MiHu, 6x KoHu, 6x OgAM, 5x HOFi, 5x KoAM, 5x DsHu, 5x MiEE, 4x CeSu, 4x MiSt, 4x MiAM, 4x MiGl, 4x DEFi, 4x HaHu, 4x HOGl, 4x HuHu, 4x OgFi, 4x HOEn, 4x HaGl, 4x TrAM, 4x HaAs, 4x MiSu, 4x HaWn, 4x HOAr, 3x OgAr, 3x KoAs, 3x KoAr, 3x OgEn, 3x CeGl, 3x HEMo, 3x DsAs, 3x HuAM, 3x TrVM, 3x MiWz, 3x HaFi, 3x
<Sequell> SpFi, 3x VpGl, 3x OpHu, 3x FoAM, 3x HuGl, 3x DsWn, 3x TrEn, 3x MuAr, 2x CeAs, 2x HOWn, 2x NaFi, 2x SEGl, 2x DEWn, 2x DsGl, 2x TeAr, 2x VSAr, 2x SpGl, 2x HOAM, 2x HOHu, 2x MfFi, 2x MDFi, 2x MiAs, 2x DDSu, 2x MuFi, 2x MuWn, 2x DsMo, 2x SpAs, 2x MDAE, 2x HuSu, 2x MiCj, 2x DrWn, 2x SpAr, 2x CeFi, 2x HaIE, 2x MuHu, 2x HaWr, 2x GrWr, 2x GrWn, 2x MiIE, 2x MiAr, 2x GrHu, 2x KoWn, 2x MiVM, CeRe, OgAs,
<Sequell> TeEn, MDWn, OgSu, MDAM, DrAs, HuWr, MuSk, DDHu, TeFi, OpAr, HOAs, MiAE, MiNe, HEAr, VSWn, HaVM, DDAM, MDGl, MDAs, SESt, FoGl, LOSk, GrFi, CeAr, DrFi, KoFi, DDAr, DsFi, GrGl, MfSk, MDWr, NaAM, HuMo, MDPa, DrSt, TrFi, DsSk, DjHu, KoGl, DEMo, HOVM, VpAs, TrHu, DESt, TrIE, OpGl, VSFi, DDWz, VSMo, MfAs, SpWn, TeAM, MiMo, KoSk, MuMo, HuWn, HaWz, FeFi, FeWn, SEHu, CeHu, DjFi, OgTm, SpHu, HuAr, DEAr,
<Sequell> DsWr, TrFE, SEWr, HuFi, OpTm, DDAs, OpFi, SpSk, LOWz, OgSk, DrWr, DDFi, OpAM, GrSu, OgMo, MfSu, DDJr, TeHu, HOAE, VSGl, TrMo, FoWr, TrWn, HuAs, HEWn, CeEn, MDEn, DjSk, TrAr

Skimming the list it appears to mostly be physical fighters, a few fighter mages such as skalds and arcane marksmen, and the very occasional weird one such as DEAr or OpTm. Most of them look like reasonable combinations.

For comparison let's check out the corresponding list for Dith.

  Code:

<halberd> !lg tenpercenters dith !zealot s=char
<Sequell> 188 games for tenpercenters (dith !zealot): 18x SpEn, 8x HaAs, 8x VSAs, 4x HEWn, 4x OpGl, 3x OpFi, 3x OpSk, 3x OpVM, 3x GrGl, 3x VSIE, 3x SpAs, 3x SpGl, 3x MuEn, 3x OpTm, 2x VSAE, 2x GrWr, 2x VSEE, 2x VSWz, 2x HaWn, 2x DsAs, 2x HaIE, 2x KoEE, 2x SpCj, 2x DsMo, 2x OgHu, 2x NaAs, 2x KoAs, 2x VpSu, 2x HuWn, 2x DsWn, 2x KoAr, 2x VpAs, 2x OpIE, 2x TeGl, 2x OpAr, 2x KoEn, 2x OpAM, 2x HaWr, 2x VpWr,
<Sequell> HEAE, TeWr, VSSk, MfCj, OgAs, VSAr, TeAr, DECj, HuEE, SpAM, GhAs, OgSu, DsCj, HuHu, MfVM, VSCj, HaHu, DsEn, DEFi, OpAE, DrAr, OpNe, MiIE, DrTm, DrMo, FoWr, GrAM, HuIE, DrCj, CeAM, KoAM, MiGl, OpHu, VpAr, HONe, TrMo, GrTm, OpCj, GrSk, DrEn, DEEn, VpEn, KoNe, HuWr, FoAE, DsNe, MiWn, FeAr, KoFi, FeCj, CeMo, OgFi, DsSk, HEAs, OpAs, VSEn, OpWz, OpWr, VSMo, SpWn, NaEn, NaAr, HuAE, NaFi, DsWr, MfEn,
<Sequell> TrFi, HOEn, VSHu

At a glance, these chars don't look noticeably less weird than the ones for trog.

[Lasty]

Thank you, and into, for a post that was as needed as it was ignored.

I think I understand the object of inquiry fairly well, being a greateroctopode.

What does winning each background of octopode have to do with understanding the relative strengths of gods, the specific strength of Trog, how to query Sequell well, or how to analyze data and come up with a clean and clear interpretation? If your point is just "I'm good at crawl, so listen to me", lots of other people in this thread are greaterthings and have better stats than you.

OpTm has the highest elo (hardest), GrFi has the lowest elo.

Well, that does a pretty good job of proving that the ELO ratings are wildly wrong.

At a glance, these chars don't look noticeably less weird than the ones for trog.

Given that you've made the case that when data is available, it should always be used, you should know that neither "at a glance" nor "noticeably less weird" pass muster. If you want to make a claim about how the output of those two queries make some kind of point, you need to give more than a shrug. Disclaimer: please don't actually try to provide a good analysis of the outputs of those two queries, since they illuminate nothing and I don't want to read any more attempts to fit an arbitrary data set into a starting conclusion.

[Berder]

Thank you, and into, for a post that was as needed as it was ignored.

I think I understand the object of inquiry fairly well, being a greateroctopode.

What does winning each background of octopode have to do with understanding the relative strengths of gods, the specific strength of Trog, how to query Sequell well, or how to analyze data and come up with a clean and clear interpretation? If your point is just "I'm good at crawl, so listen to me", lots of other people in this thread are greaterthings and have better stats than you.

No, I would not ever say "I'm good at crawl, so listen to me!" nor would I accept that from anyone else. I was responding to a paragraph where not knowing the "object of inquiry" was alluded to, with the implication that maybe I don't know the object of inquiry. The point is I'm good at crawl so it's not fair to accuse me of being unfamiliar with the data in question! It was in defense against an accusation rather than any nutty assertion that I am good at crawl therefore whatever I say is truth.

OpTm has the highest elo (hardest), GrFi has the lowest elo.

Well, that does a pretty good job of proving that the ELO ratings are wildly wrong.

What source of knowledge do you have that is better than the Elo ratings, that proves GrFi is not very easy and OpTm is not very hard? Do you know how Elo works? I think it proves that your conceptions about what is very easy and what is very hard require some adjustment. As a guess, since you didn't explain yourself, maybe you think OpTm is easy. Transmuter is the single lowest winrate background among tenpercenters, and octopode is the single lowest winrate race. Combine the two and it's not surprising OpTm has a high Elo.

At a glance, these chars don't look noticeably less weird than the ones for trog.

Given that you've made the case that when data is available, it should always be used, you should know that neither "at a glance" nor "noticeably less weird" pass muster. If you want to make a claim about how the output of those two queries make some kind of point, you need to give more than a shrug. Disclaimer: please don't actually try to provide a good analysis of the outputs of those two queries, since they illuminate nothing and I don't want to read any more attempts to fit an arbitrary data set into a starting conclusion.

I'm not the one making any assertions that trog was being used to bolster weird and weak combinations. The burden of proof for that is not on me, and a cursory glance didn't turn up anything very interesting, so I don't feel compelled to do an exhaustive analysis. As to whether it was useful data: if you believe Trog is used to bolster weak combinations, what better place to start investigating could there be than a list of combinations that took Trog?

[nilsbloodaxe]

nope, data > anecdotes.

The data we have here is actually pretty good.

First, anecdotes are a form of data, they just aren't usually all that useful. Second, we are dealing with observational data. Observational data, even in the best of cases, often fails to fit model specifications and so any inferences made may be wrong as a result of confounding. In these data, we are going to have a lot of confounding. Third, you are dealing with a non-random subset of your population (crawl games), since it only includes online games. You really can't make good inferences with non-random selection. Fourth, you are trying to do inference and model building on the same set of data (since you are trying to make a claim about the gods, while simultaneously coming up with the model of what confounding issues matter), this is a serious problem as post-selection inference often gives really bad results. Finally, even if you can say your models are good models, you don't have any indication if your measurements are both statistically and practically different.

[Arrhythmia]

No, I would not ever say "I'm good at crawl, so listen to me!" nor would I accept that from anyone else.

Why not? Perhaps naïvely, people who are good at Crawl seem to be the best people to ask why something in Crawl is good or bad.

[Berder]

First, anecdotes are a form of data, they just aren't usually all that useful. Second, we are dealing with observational data. Observational data, even in the best of cases, often fails to fit model specifications and so any inferences made may be wrog as a result of confounding. In these data, we are going to have a lot of confounding. Third, you are dealing with a non-random subset of your population (crawl games), since it only includes online games. You really can't make good inferences with non-random selection. Fourth, you are trying to do inference and model building on the same set of data (since you are trying to make a claim about the gods, while simultaneously coming up with the model of what confounding issues matter), this is a serious problem as post-selection inference often gives really bad results. Finally, even if you can say your models are good models, you don't have any indication if your measurements are both statistically and practically different.

I'm aware of these problems. I'm doing the best I can with the data I have; can you suggest a way to use the data better?

About statistical significance - yeah I'm not being that mathematical. Not that I couldn't if I took the time, but I think it's enough for these purposes to just eyeball it and estimate things like "hm 150 games, okay the sampled winrate is probably within ten percent of the true value."

No, I would not ever say "I'm good at crawl, so listen to me!" nor would I accept that from anyone else.

Why not? Perhaps naïvely, people who are good at Crawl seem to be the best people to ask why something in Crawl is good or bad.

Because personal experience is limited, and individual judgment is subjective and inaccurate, subject to cognitive biases. Also there is the effect where a strong personality claims something and others go along with it not because of evidence or even anything they personally experienced, but just to get on what looks like the winning team - a process almost entirely divorced from what's actually true. The only thing you can trust (somewhat) is the stats. They'll only lie to you to the extent you misunderstand them.

[Lasty]

No, I would not ever say "I'm good at crawl, so listen to me!" nor would I accept that from anyone else. I was responding to a paragraph where not knowing the "object of inquiry" was alluded to, with the implication that maybe I don't know the object of inquiry. The point is I'm good at crawl so it's not fair to accuse me of being unfamiliar with the data in question! It was in defense against an accusation rather than any nutty assertion that I am good at crawl therefore whatever I say is truth.

I think you misunderstood what I was saying. The object of inquiry that and into described isn't "being good at crawl", so demonstrations of crawl skill are entirely beside the point. The object of inquiry in this case is "the relative value of different gods", or something thereabouts. and into, feel free to clarify. and into's point is that critical thinking is necessary, and that simply piling up data is not helpful.

What source of knowledge do you have that is better than the Elo ratings, that proves GrFi is not very easy and OpTm is not very hard? Do you know how Elo works? I think it proves that your conceptions about what is very easy and what is very hard require some adjustment. As a guess, since you didn't explain yourself, maybe you think OpTm is easy. Transmuter is the single lowest winrate background among tenpercenters, and octopode is the single lowest winrate race. Combine the two and it's not surprising OpTm has a high Elo.

Yes, I do know how Elo works. Do you? It's used to calculate ratings in player versus player games, which you may know crawl is not. While one person attempted to shoehorn it into crawl as a way to measure player skill, doing so yielded wildly different results based on several arbitrarily-chosen variables. The suggestion that Elo should be considered a "source of knowledge" that one has to do better than is baffling. You might as well ask me to come up with a better source of knowledge than astrology or numerology.

As it happens, OpTm is not a particularly hard Op or a particularly hard Tm, and while neither Op is not the easiest species and Tm is not the easiest background, neither are they the hardest. My source of knowledge is playing a lot of crawl and having a pretty thorough understanding of crawl. Or, if it appeals to your love of data more, my source of data is running each character combo through random2(1000) ten million times and then averaging the values.

I'm not the one making any assertions that trog was being used to bolster weird and weak combinations. The burden of proof for that is not on me, and a cursory glance didn't turn up anything very interesting, so I don't feel compelled to do an exhaustive analysis. As to whether it was useful data: if you believe Trog is used to bolster weak combinations, what better place to start investigating could there be than a list of combinations that took Trog?

It's true that you're not making the argument that Trog is being used to bolster weak combinations. You're the one who needs to demonstrate that there's absolutely no confounding data and therefore all the data you offered before amounts to something other than a complete waste of time. You're also the one who doesn't seem to understand that good use of data, e.g. the kind of empirical study you're trying to emulate, requires you to make an effort to remove all confounding factors, show significance, and, if confounding factors remain that cannot be ruled out, admit them clearly and acknowledge that they may well be confounding your conclusions.

I really think it would help you to try reading some well-performed, peer-reviewed studies to see how data is correctly used in empirical research, or ideally take a class on the subject.