Fighter skill development

[comebackshane]

I'm trying to step up my game. And to this end, I'm afraid we need to use.... MATH!

So let's say I play a human fighter. Do you unwield your shield or not? what skills do you train at the very start?

I usually do something like this: keep fighting on and focus weapons skill, turn off armour and shields (but I don't unwield it). Switch to chain mail / plate mail ASAP. when I find either of them, I turn on armour as well. For shield it depends - if I get a decend two-handed weapon from a wraith pack or so, I ditch it entirely, otherwise I start turnint shield skill on when I can comfortably take out ogres and orc warriors.

Then, I try to get figthting to 10ish and weapon to min-delay while keeping armour running. I might turn everything off to throw a few points in invo/evo/throwing depending on what the dungeon gives me. I usually turn of armour at some point (15?) and start training dodging. When I keep the shield, I start training it to 15, too.

What order do you recommend? Where is the "breaking point" in str and armour xp to wear chain mail / plate mail without "overburding" yourself?

[mopl]

So let's say I play a human fighter. Do you unwield your shield or not?

If you do not plan to use the shield, you can also chose a Gladiator and start with a better weapon... Fighter have the advantage to start with better armour and !might.

About skills, I'd train only your weapon skill until it reaches min delay or near it, then some Fighting/Armour. It's usually recommended to train only one skill at a time : that's no rule, but it leads you to chose the skill you really need at each moment of the game...

What God are you planning to worship ? That can also change the way you train stuff (damage magic, buffs, trog ?)...

There is no breakpoint for str and armour, I try (rule of thumb) to have my str ~1.25 encumbrance rating.

[daggaz]

There are no breakpoints anymore, really. Also fighter is arguably not the best choice, even though it says 'fighter', for most of my melee-centered builds i will go at least with gladiator (start with helmet and nets, yes plz), but you could just as easily be a transmuter or whatnot.

Basic, very general advice is to do the following:

Turn off all skills and focus your primary weapon skill. You can choose to pump this to min_delay before anything else, but better players will just focus it until they are comfortable with their attack speed, especially for xp intensive weapons like axes and longswords.

Once you are feeling good with your weapon, either turn it off (if you hit min delay) or unfocus it and put focus in either fighting or a defensive skill, depending on your skill bonuses and your current armour type. I generally go for armour/dodging first, and get enough points there so I can see an appreciable bump in my defensive skills, but for some builds fighting is a definite good thing to do first. You can also try to push both at the same time, there is no absolute rule here.

By this time you should be approaching lair. There's a fair amount of XP in lair, and you will want to do round everything up. That means getting your current weapon to min delay, getting your armour/dodging to a level where you are comfortable in melee regarding damage absorbed, and pushing fighting as high as possible. Fighting is great, never turn it off by this point. But you will also want to put some ranks into shields (if you are using one) and possibly invo/evo depending on what you are doing.. Once again, XP is flowing and it can be quite cheap to put 7 ranks into a new skill here.

After lair, you slowly push up your 3 main skills and branch out into new skills. Orc can be useful in this, as its generally the next branch and the xp sucks, so here you turn off your high skills and focus something that is at a very low rank, instead. That way you at least get a noticeable improvement in something as you progress thru the branch, instead of taking axes from say 17.8 to 18.4.

By the later part of the game, you want a solid end-game weapon that you are swinging at min_delay or close to it (executioner's axes and claymores take a TON of xp to hit that last point, it may very well be better spent elsewhere), you want your defenses as high as possible so try to hit at least 20 in your main skill and get at least 10 ranks in the secondary, just enough to see a few points of improvement in either AC or EV, and you want fighting to be maxed. You should also have at least a few utility spells up and running, unless you worship Trog or you feel like crippling your character. Apportation, blink, swiftness, RMSL, regen, summon butterflies and the like are all very, very useful and easy to get online.

If you do extended, you just want to cap every one of these and then start pouring xp into alternative skills. You should definitely have haste castable in your plate armour.

[jejorda2]

Once I get my weapon up to 10-12 skill or so, the XP to get fighting, armour, and shields to 5-8 doesn't slow down weapon skill increase much. When weapons are around 15-20, defenses and side skills like throwing, launchers, evo, and invo cheaply get to 10-12.

Getting any skill from 10 to 20 or from 20 to 27 is harder than getting a few other skills from 0 to 10 or 10 to 15. So branch out after you get one thing working.

I tend to wear a shield until I find good armour, then take it off so I don't have too many penalties, then wear it again after I have trained armour (or shields, if I find a buckler).

[Lasty]

xp intensive weapons like axes and longswords.

Long blades are only XP intensive if you go for triple swords. Every other long blade tops out at either 14 or 16 skill. Great swords in particular are fantastic, as they're easy to find, get good brands, and do very good damage for only 16 skill -- they're very nearly as good as lajatangs, but you find dozens of them every game.

Going back to the original question, on a starting fighter, I'd keep the shield at least for the first few floors -- during that period, the shield and your decent armour combine to make for a very strong defense, even if your offense suffers a bit for it. I'd recommend turning off all skills and then turning on your main weapon skill, periodically switching it off with Fighting skill so that your Fighting skill is at about half the level of your weapon skill. If that's too micromanage-y, you could focus train the weapon and also train fighting at the same time for a 33%/66% xp split. Once you have somewhere around 8-14 weapon skill (personal taste) it's probably a good idea to bring up your other defensive skills (Armour and maybe Shields, if you plan to stick with them) to 4-8ish skill.

Overall, your weapon (or primary offensive) skill should generally be your highest skill (until it hits min delay), but you shouldn't completely neglect your other skills in the process. A rule of thumb is to try to keep your defensive skills (including Fighting) approximately at least at half the level of your primary offense. Of course that's not going to be perfect, but it's probably good enough until you get a feel for what works for you.

[Siegurt]

I have been experimenting lately with fighter starts that focus less on weapon skills and I have been training my weapons to about 5 or 6 in and fighting to about 5 or 6, and training shields up to about 9 or 10, then moving back to fighting to 10, then finally just turning all 3 of weapon fighting and shields until shields hit 15. (Often I am adding a 4th skill like evocation, dodging or armour at this point for a 25% split, depending on what items I have found)

This is actually counter to most advice, and seems to work better when paired with better body armour, but it has been working more reliably for me than the traditional 'weapon first' build. It may be something different about how I play though.

[Lasty]

I haven't tried something like Siegurt is describing in a very long time, but I have been finding that when I followed the "weapons first" strategy too closely I was dramatically under-leveling other skills to my detriment. I may still be prioritizing weapon skills too highly.

[Siegurt]

Yeah, I started experimenting with it after shields got reformed most recently, it seems to be more effective than I expected it to be.

[celem]

Personally I've never liked 'turn everything except weapon off till min'. I gather the theory behind this is that the biggest boost to your survivability in the early dungeon is actually killing stuff faster. You dont learn dodge, you just hit faster and take fewer hits in return. This only really works if you reduce each fight to a toe-to-toe tabfest 1v1.

I vastly prefer to switch on a large range of skills at the dungeon entrance and then focus the weapon (use manual training to get a sane split). I'll often have fight/dodge or fight/armour on while the weapon focuses. If it was a hybrid start I may well have spellcasting and even a spell school running too. The reason for this is I feel that in actual practical play these few points of hp/acc/ev are going to help keep me alive when im not just racing HP bars against a single foe in a corridor.

The only time this weapon-only approach is going to let me begin fighting/dodge before the point i deem 'too late' is where species apts and weapon size make it a quick sprint. But saying that, I've never actually done it, not even on a dagger.

[Sandman25]

Training more defense instead of weapon skill is an especially good idea with fighter of Chei. Of course it does not include shield, only EV and Fighting.

[mps]

I wonder if anyone could comment on the details of switching from a melee/heavy armor oriented build into a handful of good spells in the mid-to-late game. I seem to get this wrong every time I try it. Somehow I never really make the switch into using spells consistently, except song of slaying/sleighing on qaz melee orcs.

edit: particularly interested in the case of pretty bad aptitudes, e.g. minotaur, where learning a few charms and translocations seems like a real trade off.

[Sandman25]

I haven't tried something like Siegurt is describing in a very long time, but I have been finding that when I followed the "weapons first" strategy too closely I was dramatically under-leveling other skills to my detriment. I may still be prioritizing weapon skills too highly.

Probably it is caused by problems with ranged monsters like Centauri and Orc Wizards. Weapon skill is useless if you cannot attack immediately, Dodging/Armour/Fighting remain useful in this situation.

[daggaz]

Personally I've never liked 'turn everything except weapon off till min'. I gather the theory behind this is that the biggest boost to your survivability in the early dungeon is actually killing stuff faster. You dont learn dodge, you just hit faster and take fewer hits in return. This only really works if you reduce each fight to a toe-to-toe tabfest 1v1.

The early dungeon is where it is absolutely imperative to reduce each fight to a 1v1 if you want any consistent ability to survive. And yeah, the idea is correct. Avoiding damage does you no good if you cant kill at least one monster. That's not to say what I posted is the only method, tho, and I did mention that going for min_delay can be a trap.

[crate]

The scaling of offensive skills, level-for-level, is better than the scaling for defensive skills. (It's a bit tricky to get good fsim tests for this since you only get one decimal place, but as far as I can tell fsim supports this.)

I've never personally said you should train for min delay before anything else, but I would certainly say that if your defensive skills are a higher level than your offensive skills (and you're not yet past min delay etc.) then you are doing something wrong.

[Berder]

As a rule of thumb, you can figure that around 3 points of SH/EV is worth taking a 10% longer attack delay. Wearing a shield or not, for a fighter at xl1, is about the same.

[Sandman25]

I tested shield on a fighter some time ago, taking off the shield immediately is a bad idea (tested vs Adder).

[Berder]

I tested shield on a fighter some time ago, taking off the shield immediately is a bad idea (tested vs Adder).

fsim has this to say for a hufi vs an adder
with shield: aveffdam 1.1 (attacking) 0.6 (defending)
without shield: aveffdam 1.3 (attacking) 0.7(defending)
the ratio 1.3/0.7 is about the same as 1.1/0.6
therefore using the shield or not at xl1 is about equally good.

I have fsim set to use 10,000 trials (more than usual) to get greater accuracy.

I suspect your trials earlier were victims of the wily RNG.

[crate]

with shield: aveffdam 1.1 (attacking) 0.6 (defending)
without shield: aveffdam 1.3 (attacking) 0.7(defending)

To be fair, you can't really get anything from these numbers, since they're small enough that the fact you only have 1 digit after the decimal point means you're losing significant information. It could be 1.14/0.56 vs 1.26/0.74--a pretty significant difference in favor of using a shield--or it could equally likely (you can do things like run more fsim tests to figure out the relative likelihoods, or try to figure out the average yourself, but if all the information we have is a single fsim result this is equally likely) be 1.06/0.64 vs 1.34/0.66--a pretty significant difference in favor of not using a shield.

The most likely case is that they're pretty similar but the lack of precision in fsim's reporting makes a significant difference when dealing with such small numbers. It's probably actually more useful to run fsim more times and do fewer trials per attempt here, since then you can at least guess at the hidden numbers. (If you run too many trials this actually gets more difficult, since a true value of, say, 0.63 would be much more likely to show 0.6 than 0.7 in such a case--you can see this by taking the limit as you run infinitely many trials, in which case all numbers between 0.56 and 0.64 actually become indistinguishably 0.6; fewer trials can distinguish between the numbers! Since you know the number of trials you can still account for this (in theory) with any actual finite number by running more fsim tests but if you're running too many trials per fsim report it can take a while.)

This is what I was referring to in my above post.

[Berder]

You have a point about limited precision, crate. However, there's a better way to figure out what the real effect is, besides reducing the number of trials and running it many times manually. You can use the other numbers generated by fsim to calculate the shield advantage more precisely.

First we can see that the adder's accuracy when you have a shield is 50%, and with no shield it's 65%. So with no shield, you are receiving 0.65/0.50 = 1.30 times as much damage as with a shield.

Second we can see that your accuracy with a shield is 32%, and your avtime with a shield is 138. (better to use avtime than speed because it has more sig figs). And your accuracy without a shield is 35%, and your avtime without a shield is 130. So with no shield, you are dealing (0.35 / 130) / (0.32 / 138) = 1.16 times as much damage as with a shield.

So there is somewhat of an advantage to using a shield. 1.30 / 1.16 = 1.12, so you have about a 12% advantage to using it, against an adder.

Let's estimate how much uncertainty there is in that number. Running my 10,000 fsim trials multiple times, I am still getting exactly the same numbers for accuracy and avtime with and without a shield, so we can assume they are correct rounded to the last digit. First let's see what happens when we tweak the numbers as to be as favorable as possible to a shield. This means: adder's accuracy with a shield is 49.6%, adder's accuracy with no shield is 65.4%, your accuracy with a shield is 32.4%, your accuracy without a shield is 34.6%, your avtime with a shield is 137.6, your avtime without a shield is 130.4. Running the calculation again comes out to a shield advantage of 1.17 - it's unlikely to actually be this high because six numbers would have to be at the edge of their reported bounds. There's a one in ten chance for each number to be as extreme as I've tweaked them here. So, the chance of the result being this high is in the area of 1/10^6, or one in a million. So there's one in a million chance it might be as high as 1.17, and without bothering to running the numbers again we can estimate roughly equal uncertainty in the other direction, i.e. a one in a million chance it might be as low as 1.07. So that 1.12 value is fairly accurate.

The preceding analysis is quite approximate. With a more rigorous analysis the chance of a result that extreme might turn out to be as high as one in a thousand instead of one in a million. The thing to take away is that it's quite small, indicating the true value is near 1.12.

[gammafunk]

Your own fsim shows that you're not receiving 1.3 times as much damage without a shield from the adder, and I don't think that 1.3 is more accurate there due to fsim rounding. If it were closer to 1.3 you'd probably see .8 aveff damage from the adder without the shield instead of 0.7.

Also, if you modified fsim to make more trials because you care about accuracy, you should probably go ahead and modify fsim to print more digits, for the reasons crate gave.

[Berder]

Your own fsim shows that you're not receiving 1.3 times as much damage without a shield from the adder, and I don't think that 1.3 is more accurate there due to fsim rounding. If it were closer to 1.3 you'd probably see .8 aveff damage from the adder without the shield instead of 0.7.

I suggest the AvEffDam from the adder is misleading due to rounding errors. 0.74 / 0.56 = 1.32 so it's within the bounds of possibility. The number for accuracy has two digits whereas the number for the adder's aveffdam has only one digit, so it's more reliable to use the accuracy than the AvEffDam. Also, the number for accuracy doesn't depend on AvHitDam which does vary slightly due to the limited number of trials (but in theory shouldn't vary at all), which also makes the accuracy more reliable than AvEffDam, which does depend on AvHitDam.

Also, if you modified fsim to make more trials because you care about accuracy, you should probably go ahead and modify fsim to print more digits, for the reasons crate gave.

The number of trials is just a setting, fsim_rounds. see http://crawl.akrasiac.org/docs/fight_simulator.txt Figuring out how to get it to print more digits would take more effort. But if you do it I'd love to see the results.

[gammafunk]

With two decimal digits for fsim under 10,000 trials, a starting HuFi of longswords versus an adder gets, with shield:

  Code:

           AvHitDam | MaxDam | Accuracy | AvDam | AvTime | AvSpeed | AvEffDam
 Attacking:     4.86 |     12 |      33% |  1.61 |   138  |  0.72 |     1.17
 Defending:     0.87 |      4 |      50% |  0.45 |    76  |  1.32 |     0.59


and without a shield:

  Code:

           AvHitDam | MaxDam | Accuracy | AvDam | AvTime | AvSpeed | AvEffDam
 Attacking:     4.76 |     12 |      35% |  1.70 |   130  |  0.77 |     1.31
 Defending:     0.87 |      4 |      61% |  0.54 |    76  |  1.32 |     0.70


so it doesn't look like your defending accuracy was accurate without a shield, for some reason. The aveffdam from the adder is in fact close to 0.6 and 0.7 with and without a shield, respectively.

You can't disqualify use of AvHitDam simply because of variation but then not disqualify accuracy for the same reason. They both have variation from the theoretical value because they both come from a random sample.

[Berder]

That's funny; redoing the test I am still getting 65% accuracy from the adder without a shield. (Yes I'm an xl 1 HuFi with a long sword). Maybe something changed between versions. Or maybe you just got a random fluke? (Are you sure you entered 10,000 trials and not 1000?) Maybe try it again to eliminate that possibility. My local version is 0.15.2.

From your numbers the shield has an 8% advantage instead of 12%. ((0.33/138)/0.50) / ((0.35/130)/0.61) = 1.0836.

You can't disqualify use of AvHitDam simply because of variation but then not disqualify accuracy for the same reason. They both have variation from the theoretical value because they both come from a random sample.

The point is that accuracy has less variation than AvEffDam, since AvEffDam is calculated from AvHitDam, Accuracy, and AvTime, which all vary randomly, so AvEffDam would be less precise than any of its components. And only the adder's Accuracy would be affected by using a shield, so we don't need to care about the others.

[comebackshane]

I wanted to thank you for your responses, my hill orc fighters/gladiators die at least 20% less now!

[gammafunk]

I was going to make a commit for wiz-mode accuracy because of the issue that crate pointed out, but elliptic pointed out the precision for the 3rd sigdig is not very good until you get near the max limit of the fsim_rounds we support. I might do something to add this digit for aveffDam, depending on the number of rounds.

But regarding the previous discussion with Berder, aveffdam doesn't use accuracy because that's already incorporated into avdam, which will include any misses as 0. It uses two components: avdam and avtime. I did these runs with trunk, which should explain the difference with 0.15.2, possibly because of the ER changes. I think it's ideal if you just directly use aveffdam when that's the thing that you're reasoning about, but I'll have to poke around with fsim more to improve the way we display at least that number in the results. I'm also seeing some consistency issues that might be due to general fsim buginess.

[Berder]

But regarding the previous discussion with Berder, aveffdam doesn't use accuracy because that's already incorporated into avdam, which will include any misses as 0. It uses two components: avdam and avtime.

It may be that in the code AvEffDam is calculated from AvDam and AvTime, but that doesn't change its statistical dependence on the three independent components AvHitDam, accuracy, and speed.

AvEffDam = AvHitDam * accuracy * AvSpeed, and we're only interested in the accuracy component since that's the only thing affected by the shield. The other two components (AvHitDam and AvSpeed) should remain exactly constant with an unlimited number of trials; they only vary because of the limited number of trials. This variation makes AvDam less precise than the raw accuracy number.

To look at this algebraically, let D be the "true" value of AvHitDam and S be the "true" value of AvSpeed, if we had an unlimited number of trials to determine them exactly. The damage is then D*S*accuracy, and the ratio of damage between having the shield and not having it is D*S*accuracy_with_shield / D*S*accuracy_without_shield = accuracy_with_shield / accuracy_without_shield. So for the adder's damage we are only interested in the accuracy and don't care about the other two factors.

[gammafunk]

It's not very useful to reason about precision in abstract without looking at what the actual variance is causing in terms of precision, there's less math you have to do with aveffdam, and those damage values are really what you care about and would be presented anyhow to give more information than a ratio alone.The adder's accuracy is not the only thing affected by wearing a shield or that you're interested in; if it were, why would you be using avtime in the final ratio?

Debates about calculation aside, we'll improve the output from fsim in that regard in the future, hopefully for 0.16.

[Berder]

It's not very useful to reason about precision in abstract without looking at what the actual variance is causing in terms of precision, there's less math you have to do with aveffdam, and those damage values are really what you care about and would be presented anyhow to give more information than a ratio alone.The adder's accuracy is not the only thing affected by wearing a shield or that you're interested in; if it were, why would you be using avtime in the final ratio?

Accuracy is the only part of the adder's damage that's affected by wearing a shield. The player's avtime and accuracy are both affected by wearing a shield (but not his AvHitDam).

It's slightly easier just to look at the ratio of AvEffDam, but I guarantee you that if you try it many times, the ratio:

(player_AvEffDam_with_shield / adder_AvEffDam_with_shield) / (player_AvEffDam_without_shield / adder_AvEffDam_without_shield)

will have greater variance than the ratio

(Player_accuracy_with_shield * player_avtime_with_shield / adder_accuracy_with_shield) / (player_accuracy_without_shield * player_avtime_without_shield / adder_accuracy_without_shield)

and will converge to the same mean, if you increase the displayed precision enough to tell the difference and run enough tests. Therefore the second one is better.