New stats

[Magnate]

Ok, it's official - I overshot the generosity for affixes. Settling into my new career as a spin doctor, I'll tell you why this is a good thing: we've now established upper and lower bounds for item generation in v4, and can carefully tune it to be somewhere between the two. If I hadn't overshot, we'd need to keep going upwards, and would eventually overshoot and then be behind where we are now.

Owzat? Anyone going to employ me as their PR firm?

So, the details. Seven graphs attached, taken in turn:

1. Artifacts are the one thing that didn't overshoot - which is a good thing, because we know 3.3.0 was too generous. I can't work out why the curve is so much less smooth than the previous v4 curve, nor why there is no tailing off in the 90s as in the other two curves, but it's actually quite a decent shape. 1.4 artifacts per level for the last few levels, while more generous than 3.3.0, seems fair enough if the rest of the game is only half as generous. We shall see how people find it in practice (once we've fixed the over-generosity of affixes ...).

2. The "ego items per level" graph isn't terribly meaningful, but it's included here for completeness of comparison because I did it last time. This divides the total occurrence of affixes by 3 to equate with V's egos. The relevant thing here is that the number of affixes relative to the previous v4 run is 20-50% higher for most of the game, going slightly OTT past dl90. This on its own might not be a problem, but ...

3&4. The two themes graphs show where the real problem lies. Themed items are much too common now, in most cases more common than in 3.3.0. I'm going to adjust this very soon but I'm going to do it per-theme, because some of them are actually just right, while others (notably themed cloaks and high DSMs) are thousands of percent wrong.

5. Consumables are another success, now much closer to V, with the strange exception of stat potions, which seem to be more common than before despite being moved deeper. Anyone got a theory for that??

6&7. To-hit and to-dam are a mixed success: on the one hand, the number of +0,+0 items is now a mere 3x as many as V instead of 15x - this is within tolerance given the other properties these items can now have. But there are clearly far too many weapons dropping with plusses over 20, which is really going to make the game easier. These should be rarer than V. (I can only attach five files to one post so these will be separate.)

None of this is to ignore Derakon's very valid observations that missile weapons and ammo have too few affixes available, and therefore end up with way too many plusses. I'm going to try and separate these out to see whether there's still a problem with melee weapons.

And folks, could we start a new v4 observations thread again please? Anything over about 80 posts is pointlessly hard to navigate ....

[Magnate]

Placeholder for the hit/dam graphs. I'll add the ones without launchers and ammo shortly.

EDIT: ok, the problem is not restriced to launchers and ammo. If you look at the graphs which show them split out, there are still too many occurrences of high plusses for melee weapons alone - and we know that the situation for launchers and ammo is worse.

So this needs a careful re-think, to avoid being too stingy in the early game and at the same time avoid too many high plusses in the late game. (Of course, if we went to O-combat, high plusses would suffer from diminishing returns and this wouldn't be a problem .....!)

[Magnate]

Level feelings for objects

Well, some good news at last - it seems we have found the right algorithm for determining object feelings. The graph is not very smooth (need more runs), and the boundaries need adjusting (need more superb feelings), but it's the right shape: poor feelings go from top-left to bottom-right, and good ones go from bottom-left to top-right.

This is prior to the introduction of the minimum feeling for weak artifacts, btw, but it won't make much difference as there are very few of those in relation to the total number of items on a level. EDIT: the more I think about it, the more I think it works really well: the threshold curve for this minimum to apply runs smoothly from power <50 at dl1 to power <160 at dl100, which are decent definitions of "weak" for artifacts at those depths. I think even the Phial comes out with power >50 nowadays, and only about 15% of artifacts have power <160.

Thanks to Timo and others who suggested that powerful items should count for much more than weak ones.

[Timo Pietilä]

Well, some good news at last - it seems we have found the right algorithm for determining object feelings. The graph is not very smooth (need more runs), and the boundaries need adjusting (need more superb feelings), but it's the right shape: poor feelings go from top-left to bottom-right, and good ones go from bottom-left to top-right.

This is not necessarily a "right shape". Level feelings should be relative to depth, so that superb, excellent and good stay low and also bad feelings at beginning of game should be a bit lower (a single branded dagger at beginning few levels is worth "superb", but no much later). In other way around multiple uniques, few pits and couple of lesser vaults with lots and lots of low-level ego-items at clvl 98 is nothing special and isn't really "superb" or even "good".

I would have expected a lot of variance at beginning and smoothing down at "maybe" being at top for few last levels. Bad feelings and good feelings should have roughly same frequency for deeper levels as well.

[Derakon]

And folks, could we start a new v4 observations thread again please? Anything over about 80 posts is pointlessly hard to navigate ....

Have you considered using the bulletin board view instead of the threaded view? Given that the vast majority of people are presumably using the former view and just quoting who they want to respond to, I'm not certain how much utility there is in sticking with the threaded view.

Regarding the stats, I'm coming to the conclusion that we need some kind of power limiter on item generation. That is, there should be some expected power level for an item as a function of depth (or effective item level, or whatever), and the more the item exceeds that expected power, the harder it should be to generate. Unless I miss my guess, currently there's no such limiter in place; instead items are limited by the affixes that can be applied to them. But the power of a given affix can vary wildly depending on what other affixes are used -- for example, a Massive weapon is much more powerful if it also gets an acid brand.

Given how to-hit and to-dam scale item power, such an approach should rein in the outliers significantly. You could also institute minimum power levels by depth to cut down on the junk items -- though in this case, instead of trying again to generate an item of sufficient quality, you could just silently drop it.

In other words:

* Generate item.
* Calculate item power P
* Calculate minimum desired power Q, maximum desired power T
* If P < Q, destroy item
* If P > T and random((P - T) ^ 2) > fudge, go back to start and try again.
* Actually create item.

The squaring is there to make it much harder to generate an item 20 levels OOD than one 2 levels OOD, but presumably we already have a tuned algorithm to do this for us. This is just for demonstration purposes.

[Magnate]

Have you considered using the bulletin board view instead of the threaded view? Given that the vast majority of people are presumably using the former view and just quoting who they want to respond to, I'm not certain how much utility there is in sticking with the threaded view.

Hmmm. I find it helps to see how a thread has developed, i.e. which posts spawned sub-threads of conversation etc. Admittedly it's not a big deal - I just wonder why people are averse to starting new threads is all. Also, and more importantly, the threaded view makes it much easier to ignore threads I'm not interested in.

Regarding the stats, I'm coming to the conclusion that we need some kind of power limiter on item generation. That is, there should be some expected power level for an item as a function of depth (or effective item level, or whatever), and the more the item exceeds that expected power, the harder it should be to generate. Unless I miss my guess, currently there's no such limiter in place; instead items are limited by the affixes that can be applied to them. But the power of a given affix can vary wildly depending on what other affixes are used -- for example, a Massive weapon is much more powerful if it also gets an acid brand.

The power algorithm is designed to recognise precisely this phenomenon: it calculates expected damage, and therefore calculates exactly how effective different damage-affecting attributes are in combination. It's not quite as specific for non-damage attributes, but it does have additional terms for multiple resists, sustains, pval-related abilities etc. etc.

The problem with expected minima and maxima is that they require the use of arbitrary constants, which I am dedicating my efforts to avoid. Otherwise every time I amend the power algorithm all the arbitrary constants for power thresholds elsewhere in the code need amending. At the moment the only ones outside obj-power.c are the level feelings, and I'd rather not use more.

Instead, object power is limited by inhibiting certain qualities (e.g. +3 blows etc.). These are of course absolute limits and therefore not relative to depth - I have to go and put the children to bed now but I'll do some more thinking about this. So far I've been quite happy with the idea that limiting the number of affixes by depth would self-limit the power of items - I think if we're not too generous that will create the occasional "oh wow" moment that we all like without being unbalancing. But I could be wrong ...

[fizzix]

-- for example, a Massive weapon is much more powerful if it also gets an acid brand.

This is also the reason why having accurate power levels is so incredibly hard to generate. massive + acid is synergistic and would provide a boost. Similarly a cold branded weapon that gets slay undead has added utility, but a fire branded weapon that gets slay undead does not gain much added utility beyond helping to off some ringwraiths. There are ways to account for this synergy but it's very hard to get right, and I'm not sure it's worth it.

[Derakon]

Couldn't you calculate the damage per blow for each monster, multiply that by the monster's power level, and come up with a weighted average melee power? You'd also probably want to weight the average by the weapon's weight to reflect the potential for extra blows. That Massive Great Hammer of Disruption I found that weighs in at 90 pounds would be a notable example -- I only got 3.3 blows with it and frankly it's astonishing I did that well...

EDIT: as for why I don't create new threads: given that my posts tend to be random collections of observations, and that they tend to spawn short-lived conversations about random topics, I'd be worried about thread spam, not to mention being unable to come up with meaningful thread titles.

[Magnate]

Couldn't you calculate the damage per blow for each monster, multiply that by the monster's power level, and come up with a weighted average melee power?

Gents, this is exactly how it already works! Chris Robertson thought of this eleven years ago and the code still survives in obj-power.c (function slay_power).

EDIT: as for why I don't create new threads: given that my posts tend to be random collections of observations, and that they tend to spawn short-lived conversations about random topics, I'd be worried about thread spam, not to mention being unable to come up with meaningful thread titles.

Someone came up with a solution to that a while ago: you create something called "Derakon's thread" and add all your random observations to that (where they don't fit neatly into other more specific threads). ;-)