Team Robot Simulator and Gear Comparison Tool - Page 9

Yellowsix · 01/29/10, 6:02 AM

I posted a quick update with some bug fixes. Here's the patch notes:

- fixed an issue where items with on-use effects would sometimes cause rotation logic to get "stuck"

- improved estimates for nibelung: the combat log may show higher average damage per hit than is expected, this is because concurrent val'kyr summons are not modeled individually -- they are combined into one and treated sort of like a "rolling DoT"; the total damage is correct though

- added ashen verdict caster ring proc

- fixed a timing bug that was causing cooldown checks to happen slightly too early (a good example was solar eclipse activating just slightly before lunar eclipse cooled down)

Yellowsix · 01/30/10, 6:49 AM

Posted another update with some minor tweaks to the default feral rotation.

A few thoughts on the feral rotation:

The current Mr. Robot default is a bit different than the SimC rotations that are floating around. For those of you who want to try the SimC rotation that Furion posted earlier in this thread in the Mr. Robot simulator, you can download it from http://www.askmrrobot.com/xml/feral_rotation.xml and import it (see WoW Simulator Import/Export Feature for instructions).

A couple things to note:

1. "time to live" conditions are currently not in either the Mr. Robot default or the linked transcription of the SimC rotation. I'll be improving a couple of the conditions in Mr. Robot to make this easier to do, but it's not ready yet. For feral it is a very marginal difference in DPS though, and would not affect gearing decisions either. Those also don't translate into actual game play very well either, since time to live can be hard to predict in a real fight, particularly at the short intervals of use to feral.

2. The Mr. Robot default rotation does more damage than the SimC rotation. I tried many combinations, and I just think it's a better rotation -- either that or I didn't transcribe the SimC rotation correctly, which is entirely possible. If anyone can beat the current default, please let me know, and send me the XML for your rotation.

3. The Mr. Robot default rotation is intuitive and practical. One could execute this rotation without a move-suggesting mod, though it borders on getting too complex. In constrast, the SimC rotation is purely academic. Even if someone could tweak it and squeeze out another 2% DPS, I'd be inclined to stick with a simpler default rotation, and make the complex one available as a separate pre-defined rotation for theorycrafters to toy with.

4. Some of the parameters and conditions in the feral rotation may need tweaking depending on your gear and set bonuses. For instance, you might aim for higher rake up-time with one of the rake set bonuses, or if you have low armor penetration you might aim for higher rip/rake up-time and less FBs. There are definite tipping points. Play around with it! All in all though, the default should be pretty reasonable with any gear.

I know that a lot of you have been tweaking feral rotations for a long time, but I think that the Mr. Robot rotation and simulator is worth a look.

Furion · 01/30/10, 7:27 AM

Well you definetly want to shred when berserk is up and there is nothing else to do. This also seems to increase dps in your simulator. Also with 4T10 you will definetly want a higher rake uptime. Also the DPS variance between simulations makes testing kind of hard: when I removed the "use bloodfury" condition (since I am not orc) my dps increased by 100 (Edit: Ok I just increased the number of runs to 1000 which will decrease the dps variance)
Also you might want to include time spent at energy cap in your summary.

Allev · 01/30/10, 12:09 PM

Originally Posted by Yellowsix

Posted another update with some minor tweaks to the default feral rotation.

A few thoughts on the feral rotation:

The current Mr. Robot default is a bit different than the SimC rotations that are floating around. For those of you who want to try the SimC rotation that Furion posted earlier in this thread in the Mr. Robot simulator, you can download it from http://www.askmrrobot.com/xml/feral_rotation.xml and import it (see WoW Simulator Import/Export Feature for instructions).

A couple things to note:

1. "time to live" conditions are currently not in either the Mr. Robot default or the linked transcription of the SimC rotation. I'll be improving a couple of the conditions in Mr. Robot to make this easier to do, but it's not ready yet. For feral it is a very marginal difference in DPS though, and would not affect gearing decisions either. Those also don't translate into actual game play very well either, since time to live can be hard to predict in a real fight, particularly at the short intervals of use to feral.

2. The Mr. Robot default rotation does more damage than the SimC rotation. I tried many combinations, and I just think it's a better rotation -- either that or I didn't transcribe the SimC rotation correctly, which is entirely possible. If anyone can beat the current default, please let me know, and send me the XML for your rotation.

3. The Mr. Robot default rotation is intuitive and practical. One could execute this rotation without a move-suggesting mod, though it borders on getting too complex. In constrast, the SimC rotation is purely academic. Even if someone could tweak it and squeeze out another 2% DPS, I'd be inclined to stick with a simpler default rotation, and make the complex one available as a separate pre-defined rotation for theorycrafters to toy with.

4. Some of the parameters and conditions in the feral rotation may need tweaking depending on your gear and set bonuses. For instance, you might aim for higher rake up-time with one of the rake set bonuses, or if you have low armor penetration you might aim for higher rip/rake up-time and less FBs. There are definite tipping points. Play around with it! All in all though, the default should be pretty reasonable with any gear.

I know that a lot of you have been tweaking feral rotations for a long time, but I think that the Mr. Robot rotation and simulator is worth a look.

Sadly, I can't actually load anything from the armory right now and I'm on a new machine, so Mr. Robot is unusable for me right now. It would be valuable to at least include a "default" feral/moonkin profile like you do for DKs. But, let me address some of your points:

1. Without a TTL, you'll have trouble making a fair comparison. The SC profile assumes that you'll burn excess energy at the end of the fight. Without doing this, you're using less total energy and losing a few attacks.

2. Can you quantify this amount? Are you talking 10 DPS or 300 DPS? Again, I can't see the Robot cycle right now. But, perhaps it'll be worthwhile to also implement your cycle in SC, as it may enlighten us as far as bug-hunting in both sets of code. Also, describing the actual differences would be valuable.

3. I've never used a move-suggesting mod, and the SC profile is certainly not that difficult. The decisions made mostly involve shredding in certain situations, and not shredding in others. It's not "purely academic" any more than Mr. Robot's would be. The actual problem is the SC profile is actually incomplete, in that it doesn't make every decision that an experienced feral will make-- part of that is because they are difficult to narrowly formulaically specify, part of that is because those decisions are small/rare enough that they aren't statistically significant.

4. Are you testing in players' actual gear, or BiS gear?

Furion · 01/30/10, 12:32 PM

For my character (only 2 T10 unfortunately) it shows around 300 dps lost by using the sc action priority list although the simcraft profile has slightly higher uptimes on rip and SR and significantly higher uptime on rake.
For starting out I probably should have suggested a simpler version of the sc action priority list as not every detail is as important as others. I think the SR refreshing part isn't translated correctly and this one is very important.

actions+=/savage_roar,cp>=3,if=buff.savage_roar.remains<=8&dot.rip.remains-buff.savage_roar.remains>=-3

is translated as "difference in time remaining on roar and rip is less than 3 seconds" But I think it needs to be vice versa rip needs to be first (makes about 200 dps difference from testing)

Edit: I translated the move suggestions to simcraft and it resulted in a huge (600 dps) loss with BiS gear (if my translation is correct) vs the action priority list I currently use in SC. I attribute this mainly to only 73% rake and 75% rip uptime in simcraft, while team robot simulator assumes 87% for rip and 76% for rake. For my standard SC profile uptimes of rip+rake are around 90% in both simulators.

actions+=/tigers_fury,energy<=30,berserk=0
actions+=/berserk_cat,energy>=70,energy<=90,if=buff.tigers_fury.down
actions+=/savage_roar,cp>=2,if=buff.savage_roar.down
actions+=/rip,cp>=5
actions+=/savage_roar,cp>=5,if=buff.savage_roar.remains<4&!dot.rip.remains<4
actions+=/ferocious_bite,cp>=5,if=buff.savage_roar.up&!dot.rip.remains<6
actions+=/shred,extend_rip=1
actions+=/shred,if=buff.omen_of_clarity.up
actions+=/rake,if=buff.combo_points.stack<=4|energy>=90
actions+=/shred,if=(buff.combo_points.stack<=4|energy>=90)

Yellowsix · 01/30/10, 3:53 PM

Note that abilities in the Mr. Robot rotation like blood fury or trinket uses or heroism will not trigger unless you have the required gear, race, or raid buffs. The default just puts a bunch in there as a convenience -- they'll execute only when appropriate.

Another thing to note -- I don't have haste potions implemented yet, so I left that out of both rotations.

I realize that putting time-to-live conditions in will definitely increase DPS. I left them out of both rotations for now, so the comparison between the two should still be valid.

It seems that there are definitely some differences in the decision engines used by the Mr. Robot simulator and SimC. I find it interesting that trying to do the SimC rotation in Mr. Robot loses about 300 DPS, but doing the Mr. Robot rotation in SimC loses about 600 DPS. Would be nice to figure out what each simulator is doing differently.

I did these tests on my current gear, which is not optimal. I'll try making a rotation that uses 4T10 and BiS gear, see if a different priority is optimal in that case. As you have suggested, I would expect favoring higher rake up-time to be desirable with 4T10.

As for rip up-time, I find that there is a trade-off between rip and the number of ferocious bites that you use. This tipping point does not seem to be constant for all ferals, and probably needs to be tweaked to the individual player to get optimal DPS. Letting rip drop here and there isn't that huge a deal, as long as it's not for too long. One ferocious bite is usually worth about 4-5 ticks of rip, so if you're losing less than that to get off an FB, you're coming out ahead.

Yellowsix · 01/30/10, 5:32 PM

I did a comparison between the SimC rotation and the Mr. Robot rotation using both simulators and my character (Yellowsix). I don't have optimal gear of course -- I'll do a BiS simulation later today when I have time. But here's the results:

(I did 1000 iterations of 300 second fights, with gcd_lag set to 50ms in both simulators)

I removed all "time to live" conditions from all cases. I also removed potions.

SimC:
SimC rotation: 9211 DPS
Mr. Robot rotation: 9093 DPS

Mr. Robot:
SimC rotation: 8910 DPS
Mr. Robot rotation: 9035 DPS

It seems that there is some amount of "gaming" the simulator that can be done to produce optimal results with both simulators: the SimC rotation works better with SimC, and the Mr. Robot rotation works better with Mr. Robot. Nevertheless, the results are all pretty close considering that the code and design of each simulator is completely independent.

I'll do a BiS comparison later, which I assume will require a slightly tweaked rotation.

Allev · 01/30/10, 6:21 PM

Do you implement any kind of energy pooling? If you don't, then you'll never have over 42 energy (except right after using TF). Whereas the SC profile often ends up over this value, frequently up to 70 or higher with OoC procs. The TTL codes are to make sure you don't have energy you could have used at the very end of the fight. If you get one more 10k shred in, you get another 33 DPS for the fight. The real difference is probably 20 DPS on average.

I haven't personally seen a large difference between T10 BiS and everything else-- the numbers slightly change when going from soft ArP cap to hard ArP cap (from ~6 seconds left on Rip to ~8-10 seconds left on Rip) but that's less about T10 and more about the ArP. Before 4T10, shred can overcome Rake at ArP cap, eliminating it from the sequence.

Yellowsix · 01/30/10, 9:11 PM

I don't think that we're quite communicating Allev...

Firstly, if you run a SimC simulation or a Mr. Robot simulation, using either of the rotations posted above, you will see a roughly similar breakdown of ability usage, and a roughly similar DPS output. If one simulation or the other were not implementing "energy pooling" as you say and allowing 42 energy to build up for shred, you could be sure that the differences would be a lot larger than +/- 100 DPS.

As far as I know, neither SimC nor Mr. Robot implement special "energy pooling" explicitly (correct me if I'm wrong) -- it's unnecessary. If you feed a priority-based simulator the proper priority logic, it will decide to wait until it meets the correct conditions to perform an action, one of those conditions being "do I have enough energy?"

Secondly, I have stated that I'm going to be adding TTL conditions in an upcoming update, because they do indeed increase DPS slightly. But only slightly. 1 extra FB in a 5-minute fight is hardly worth getting excited about. But since these conditions are not ready yet, when comparing SimC and Mr. Robot rotations, I have removed them. This will yield slightly less than "optimal" DPS, but it will yield a valid direct comparison of the two simulators.

And yes, at some point it is possible that it becomes inefficient to use Rake at all from a DPE perspective, if your ArP to AP ratio is high enough, and you don't have any of the Rake-enhancing set bonuses.

I'll be comparing SimC and Mr. Robot more extensively over the weekend, and see if I can discover why the Mr. Robot rotation performs almost the same in both simulators, but the SimC rotation swings 300 DPS between the two.

Allev · 01/31/10, 2:57 AM

So, I got Mr. Robot to load here.
- It's really hard to see what's a DPS increase and what isn't when two simulation sessions with 1000 iterations end up varying by up to 100 DPS (well outside of expected error).
- You shred to 5 CP, constantly. This will lead to dropping Rake uptime a lot and dedicating more CP to FB than it's really optimal to do.
- The strategy of always shredding when not at 5 CP means that you're rarely conserving energy-- essentially, only when you're waiting for Rip to expire.
- You aren't even getting that many additional FBs-- loading my character, I only get 2.4% damage from my FBs.
- You may attack twice without an SR buff if you don't crit the first one. That's just a bad idea.

In other words, you're doing tons of things in your cycle that are logically sub-optimal. My guess is that the SC profile didn't get trans-coded properly to Mr. Robot, if all mechanics are functioning properly in each.

Edit: Something is wrong with the "Target missing SR or less than X seconds remain". With X = 8, a 4 CP SR is being refreshed after ~14 seconds. Either your SR length is computed wrong or the phrase simply isn't doing what it's designed to do. Is it triggering the "missing SR" after the 1-point SR has expired?

Yellowsix · 01/31/10, 5:58 PM

- It's really hard to see what's a DPS increase and what isn't when two simulation sessions with 1000 iterations end up varying by up to 100 DPS (well outside of expected error).

When I've been running at 1000 iterations with Yellowsix, I haven't seen the result fluctuating by more than 10-20 DPS in either direction, which is within the estimated margin of error... I don't claim to be a statistics wizard though. I'll try it with some different characters and cases, but as of now I cannot reproduce this behavior. I'll keep an eye on it.

- You shred to 5 CP, constantly. This will lead to dropping Rake uptime a lot and dedicating more CP to FB than it's really optimal to do.

This is just an assertion... I can't really do anything with this statement. So say I lose 10% rake uptime, which is about 10 ticks in a 5-minute fight. Why is it obvious that this is a problem? There are too many variables to just assert this as truth. Same for dedicating more CPs to FB... just saying that it is not optimal does not make it so... thus simulators! It would be better if you were to create a rotation with higher rake up-time and less FBs that does more damage, and show it to me.

- You may attack twice without an SR buff if you don't crit the first one. That's just a bad idea.

Same as above -- tell me why, or better yet, show me why with a custom rotation -- don't just assert that it is true, or I can't do anything with it. Let me try and think about this particular statement though... let's estimate this potentially bad scenario: I get 4 auto-attacks, 2 rip ticks, 1 rake tick, and 2 shreds without SR up. With Yellowsix, I would have done ~13,000 extra total damage had SR been up. So then the question becomes... what did I gain by allowing that gap in SR up-time? Say I gained one FB. My average FB damage is ~19,500. Thus, I gain 6,500 damage overall.

While I agree that gaps of SR downtime is probably a bad thing... it is not trivial to determine at what point 100% up-time is worth potentially sacrificing an extra FB during the fight.

Something is wrong with the "Target missing SR or less than X seconds remain". With X = 8, a 4 CP SR is being refreshed after ~14 seconds. Either your SR length is computed wrong or the phrase simply isn't doing what it's designed to do. Is it triggering the "missing SR" after the 1-point SR has expired?

I checked the code, and it was possible that an SR could be executed as if the player had less CPs than they actually had. I have fixed this, so this situation shouldn't happen anymore. Good catch.

Yellowsix · 01/31/10, 6:11 PM

I have taken a few of the suggestions into account and done another comparison to SimC (using Yellowsix, still have to build a BiS case).

The new Mr. Robot rotation is posted, and here is the SimC version of it:

actions=flask,type=endless_rage
actions+=/food,type=hearty_rhino
actions+=/cat_form
actions+=/auto_attack
actions+=/snapshot_stats
actions+=/faerie_fire_feral,debuff_only=1
actions+=/tigers_fury,energy<=30,berserk=0
actions+=/berserk_cat,energy>=70,energy<=90,if=buff.tigers_fury.down
actions+=/savage_roar,cp>=2,if=buff.savage_roar.down
actions+=/savage_roar,cp>=5,if=buff.savage_roar.remains<4&dot.rip.remains>=4
actions+=/rip,cp>=5
actions+=/ferocious_bite,cp>=5,if=buff.savage_roar.up&dot.rip.remains>=6
actions+=/shred,extend_rip=1
actions+=/shred,if=buff.omen_of_clarity.up
actions+=/rake,if=buff.combo_points.stack<=4|energy>=90
actions+=/shred,if=((buff.combo_points.stack<=4&dot.rake.remains>=4)|energy>=90|buff.berserk.up)

The suggestions that I incorporated were changes to the shred condition to use it when berserk is up and there's nothing else to do, and also to keep higher rake up-time.

Also, I properly implemented the following condition in SimC into the Mr. Robot simulator:

actions+=/savage_roar,cp>=3,if=buff.savage_roar.remains<=8&dot.rip.remains-buff.savage_roar.remains>=-3

Before it wasn't doing this quite correctly.

Results:

SimC
SimC rotation: 9229 DPS
Mr. Robot rotation: 9173 DPS

Mr. Robot
SimC rotation: 8901 DPS
Mr. Robot rotation: 9325 DPS

So as before... there's something about the SimC rotation that is not translating into the Mr. Robot simulator quite right. I think that either the feral code or the energy-tracking code in SimC must be doing some behind-the-scenes magic. For example, you don't have to code into the rotation logic "don't apply a DoT if it's already up." This is not generally true for all classes, so I'm assuming that it's part of the feral code, not SimC as a whole. Also, SimC seems to be doing some kind of behind-the-scenes energy conservation in order to build up for berserk... Mr. Robot will not implicitly do this. I have lowered the energy threshold to 70 for the SimC rotation as implemented in Mr. Robot, which causes it to use berserk twice for me. Not ideal, but works for now.

Also, SimC does not seem to be modeling the on-use effect of Victor's Call, or else it has to be explicitly put into the rotation somehow, but I'm not sure how to do that.

dedmonwakeen · 01/31/10, 7:16 PM

First things first: Just because SimC was born before MrRobot does not mean SimC is "more" right.

The default behavior for DoTs is to NOT let them get overwritten..... however, with the introduction of the if= conditionals that is something in the process of being changed.

I would examine damage break-downs. How many of each type of ability is used? (on average, fractional values)

I would examine the rate of CP generation (and CPs "wasted" due to 5-stack).

I would also closely examine the rate of energy regen and how each sim "waits" for energy regen. SimC performs power-regen every 0.1sec by default. These are particular/arbitrary modeling decisions that could potentially have an unexpectedly large impact on DPS. When a player comes to the end of his action list with nothing to do, he "waits" for some period of time. I used to make this a super-tiny value but it was killing my performance. Now it is a function of current energy levels.

This is the function used to determine how far into the future to insert the next "player_ready" event when the player comes to the end of his action list.

double druid_t::available()
{
  if( primary_resource() != RESOURCE_ENERGY ) return 0.1;

  double energy = resource_current[ RESOURCE_ENERGY ];

  if ( energy > 25 ) return 0.1;

  return std::max( ( 25 - energy ) / energy_regen_per_second, 0.1 );
}

Yellowsix · 02/01/10, 1:39 AM

I'll add something to what dedmonwakeen said: just because Mr. Robot is new and shiny doesn't mean that it's trying to replace SimC or prove that SimC is wrong.

I think that the major takeaway from these comparisons of SimC and Mr. Robot is that both simulators are very close. Sure there are a few minor things to dig into and figure out, mainly with some probably minor differences in how each simulator decides which ability to use next, but on the whole, two independently-developed simulators are giving confirming results. This is a good thing for both simulators.

I ran into similar design issues dedmonwakeen: how long should I wait before checking again if I can't perform an action? I will double check that Mr. Robot isn't waiting too long to re-evaluate the condition logic. Whereas you had the issue of checking too often, I ran into the problem of not checking often enough and fixed it, but it's possible that I missed a case or two.

Allev · 02/01/10, 5:24 AM

Originally Posted by Yellowsix

When I've been running at 1000 iterations with Yellowsix, I haven't seen the result fluctuating by more than 10-20 DPS in either direction, which is within the estimated margin of error... I don't claim to be a statistics wizard though. I'll try it with some different characters and cases, but as of now I cannot reproduce this behavior. I'll keep an eye on it.

This is just an assertion... I can't really do anything with this statement. So say I lose 10% rake uptime, which is about 10 ticks in a 5-minute fight. Why is it obvious that this is a problem? There are too many variables to just assert this as truth. Same for dedicating more CPs to FB... just saying that it is not optimal does not make it so... thus simulators! It would be better if you were to create a rotation with higher rake up-time and less FBs that does more damage, and show it to me.

Same as above -- tell me why, or better yet, show me why with a custom rotation -- don't just assert that it is true, or I can't do anything with it. Let me try and think about this particular statement though... let's estimate this potentially bad scenario: I get 4 auto-attacks, 2 rip ticks, 1 rake tick, and 2 shreds without SR up. With Yellowsix, I would have done ~13,000 extra total damage had SR been up. So then the question becomes... what did I gain by allowing that gap in SR up-time? Say I gained one FB. My average FB damage is ~19,500. Thus, I gain 6,500 damage overall.

While I agree that gaps of SR downtime is probably a bad thing... it is not trivial to determine at what point 100% up-time is worth potentially sacrificing an extra FB during the fight.

I checked the code, and it was possible that an SR could be executed as if the player had less CPs than they actually had. I have fixed this, so this situation shouldn't happen anymore. Good catch.

I've always had the fluctuating DPS problem-- maybe it's something specific with the gear my character is currently using (NES/DMC:G/Idol of Mutilation/Berserking enchant are the odd items).

Trading a higher DPE attack (i.e. Rake) for a lower DPE attack (i.e. shred) is clearly a good thing. Waiting for Rakes instead of shredding all the way until 5 CP is increasing your DPE, while simultaneously creating more CP than if you shredded (since it is less energy per CP).

The requiring 2 points to cast SR (in every circumstance) simply means that if SR falls, you can attack twice (instead of once) when starting your cycle. You aren't generating 5 extra CP by delaying using your SR. I'm really trying hard, and failing, to understand your logic on how this increases your DPS. I don't see any basis for a DPS increase unless you are trying to stabilize a healthy rotation; but you're throwing away roughly 1/4 of each attack by not having SR up already.

The building up energy in the SimC profile is completely done by not shredding unless you have a reason to do so. Shredding when berserk is up is one of those reasons; OoC, high energy, TF coming off cooldown, or needing more CP for a finisher are all good reasons. And this naturally gets us above thresholds to use Berserk and below thresholds to use TF.

I don't see MrRobot as a threat to SimC, and in fact I look forward to the time where your sim is as tested and robust as SimC is. But right now, I don't think it is approachable as a finished product when you're still uncovering fundamental mechanical bugs.