GA statistical musings
Moderators: wdolson, MOD_War-in-the-Pacific-Admirals-Edition
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
GA statistical musings
So, I flipped some turns in some scenario to look onto some data the game would generate. Some results were not what I would expect.
Scenario: Coral Sea stripped of all Allied activity and loaded with all things Japanese, starting with Dec-41. Scenario files attached for the curious.
Batch of tests #1: Aircraft R&D
600 factories of size 30 (100 per airframe), stuffed into New Guinea/Britain and Solomons bases with ample supply. 4- or 1-day turns depending on how far away the airframe arrival was (turned out it did not matter). About 500 turns flipped by starting scenario and doing one cycle, reading data then starting scenario again.
Main question being how the game repairs the factories. I.e. what is the game formula for the probability of repairing 1 unit during a turn.
A guess that was circulating the forums so far (from here) is that the probability P=min[1;S/T] where S is factory size and T is time in days left until airframe arrival. Tests show this guess is roughly OK, but with one significant quirk.
Game calculates stuff in months not days, so the probability of repair is constant during a calendar month. E.g. a 30-size factory with less than a month left until arrival will have exactly 100% chance of repair each turn. But if 31-59 days are left, the chance is ~56,1%. If 60-89 days the chance is ~35,5%. Et cetera.
Good approximation for the probability of repair for 30-size factory would be P=min[1;30/(months*30-8)] where "months" is the number of months to arrival rounded up. Was too lazy to look in depth on other sizes of factories, as 30 is optimal for R&D. But the formula is definitely multiplicative in terms of factory size. Which is ofc not correct wrt true probabilistic behaviour, but that is how game is programmed. E.g. in game two 15-size factories will not repair the same way as one 30-size.
Batch of tests #2: off-base fort building
This thread have started the itch, with the main question being why the heck IDs can't build forts higher than level 2. So I put a bunch of test divisions with different TOEs/commanders/loading costs of devices. At later stages I settled for the modified TOE (450 inf squads + varying amount of support/motosupport and nothing else) to catch the possible loading cost or device number thresholds. It turns out that size of a unit seem to play a major part in deciding if forts can go higher, and looks like size is approximated by loading cost.
Main finding: LCUs of load costs <8000 were able to build forts higher than 3, where LCUs of load cost >11-12k were consistently stuck with level 1-2. All Japanese brigades and regiments fall into the first category, while most of IDs fall into the second, unless you divide them into A/B/C regiments.
I cound not determine what drives IDs to be stuck at forts 1 sometimes.Also, I played only with Japanese LCUs because that's where it is most important. Chinese can dig away just fine already, and other allies do not need off-base fort building that much.
Main finding #2 (surprising) is that disablements do not matter for fort building speed.
Batch of tests #3: morale and experience buildup
Dozens of 30/30 XP/morale divisions with 100% prep sitting in place for 60 turns in combat mode. XP increased to 40-52. Done a spreadsheet table of all the relevant numbers for couple dozen IDs.
Main finding #1 (surprising) is that Leadership skill of ID commander plays absolutely no role in how XP increases. Other skills also do not correlate with XP increase.
Main finding #2 (not surprising) is that Inspiration skill helps increase morale faster. After 60 turns IDs increased morale from 30 to 70-98, with inspiration 40-50 sticking in the lowest part, max inspiration of 77 getting morale to the max too (98), and generally visible correlation between higher insiration and higher end morale
Batch of tests #4: rest and recuperation
Dozens of 100% disabled divisions in base and non-bases hexes in rest mode, combined with lots of support units (HQs with lots of support squads) in combat mode.
Main finding #1 (surprising) is that Admin skill of Div commander plays absolutely no role in how quickly squads would repair. Other skills also do not correlate with repair speed.
Main finding #2 (a bit surprising) is that IDs seem to repair slightly faster in a developed base hex compared to non-base, but not that much faster. In 60 turns standard TOE IDs repaired about 23-30% in base and 16-25% in non-base hexes.
Main finding #3 (not surprising) is that ABC-divided IDs repair significantly (~2 times) faster compared to whole IDs.
Batch of tests #5: effects of LCU morale on mundane stuff
As per some of Lowpe's questions here tm.asp?m=4265621
80 LCUs of different variety sit (or marched) for several turns with either morale 30 or morale 99 (changed for all of them in scenario editor between tries), all other things being equal.
Main finding Morale does not affect: engineers and avsupport efficiency, marching speed, fatigue/disruption/disablements recovery speed.
Batch of tests #6: effects of HQs
Main finding #1 (not surprising) Presense of HQ/HQc/HQy and the fact of preparation gives significant bonus to LCUs attacking the base. LCU without HQ in range perform at 50% efficiency, that is adjusted AV will be halved on average compared to base AV. HQ in range increases efficiency to 70% even w/o prepping for the base. HQ prepping for the base can increase efficiency up to average 150% (with all units 100 prepped), and up to double in single attacks.
Main finding #2 (mildly surprising) Preparation of attacking LCUs to the same base is necessary to use HQ prepping bonus. Average efficiency increase linearly with LCU prep level (from ~70 to ~150) due to HQ bonus. LCU prep level w/o HQs does not matter much for AV.
If you have any suggestions for other tests, I might be able to run them too. No airgroups though, those are a whole new hassle.
Scenario: Coral Sea stripped of all Allied activity and loaded with all things Japanese, starting with Dec-41. Scenario files attached for the curious.
Batch of tests #1: Aircraft R&D
600 factories of size 30 (100 per airframe), stuffed into New Guinea/Britain and Solomons bases with ample supply. 4- or 1-day turns depending on how far away the airframe arrival was (turned out it did not matter). About 500 turns flipped by starting scenario and doing one cycle, reading data then starting scenario again.
Main question being how the game repairs the factories. I.e. what is the game formula for the probability of repairing 1 unit during a turn.
A guess that was circulating the forums so far (from here) is that the probability P=min[1;S/T] where S is factory size and T is time in days left until airframe arrival. Tests show this guess is roughly OK, but with one significant quirk.
Game calculates stuff in months not days, so the probability of repair is constant during a calendar month. E.g. a 30-size factory with less than a month left until arrival will have exactly 100% chance of repair each turn. But if 31-59 days are left, the chance is ~56,1%. If 60-89 days the chance is ~35,5%. Et cetera.
Good approximation for the probability of repair for 30-size factory would be P=min[1;30/(months*30-8)] where "months" is the number of months to arrival rounded up. Was too lazy to look in depth on other sizes of factories, as 30 is optimal for R&D. But the formula is definitely multiplicative in terms of factory size. Which is ofc not correct wrt true probabilistic behaviour, but that is how game is programmed. E.g. in game two 15-size factories will not repair the same way as one 30-size.
Batch of tests #2: off-base fort building
This thread have started the itch, with the main question being why the heck IDs can't build forts higher than level 2. So I put a bunch of test divisions with different TOEs/commanders/loading costs of devices. At later stages I settled for the modified TOE (450 inf squads + varying amount of support/motosupport and nothing else) to catch the possible loading cost or device number thresholds. It turns out that size of a unit seem to play a major part in deciding if forts can go higher, and looks like size is approximated by loading cost.
Main finding: LCUs of load costs <8000 were able to build forts higher than 3, where LCUs of load cost >11-12k were consistently stuck with level 1-2. All Japanese brigades and regiments fall into the first category, while most of IDs fall into the second, unless you divide them into A/B/C regiments.
I cound not determine what drives IDs to be stuck at forts 1 sometimes.Also, I played only with Japanese LCUs because that's where it is most important. Chinese can dig away just fine already, and other allies do not need off-base fort building that much.
Main finding #2 (surprising) is that disablements do not matter for fort building speed.
Batch of tests #3: morale and experience buildup
Dozens of 30/30 XP/morale divisions with 100% prep sitting in place for 60 turns in combat mode. XP increased to 40-52. Done a spreadsheet table of all the relevant numbers for couple dozen IDs.
Main finding #1 (surprising) is that Leadership skill of ID commander plays absolutely no role in how XP increases. Other skills also do not correlate with XP increase.
Main finding #2 (not surprising) is that Inspiration skill helps increase morale faster. After 60 turns IDs increased morale from 30 to 70-98, with inspiration 40-50 sticking in the lowest part, max inspiration of 77 getting morale to the max too (98), and generally visible correlation between higher insiration and higher end morale
Batch of tests #4: rest and recuperation
Dozens of 100% disabled divisions in base and non-bases hexes in rest mode, combined with lots of support units (HQs with lots of support squads) in combat mode.
Main finding #1 (surprising) is that Admin skill of Div commander plays absolutely no role in how quickly squads would repair. Other skills also do not correlate with repair speed.
Main finding #2 (a bit surprising) is that IDs seem to repair slightly faster in a developed base hex compared to non-base, but not that much faster. In 60 turns standard TOE IDs repaired about 23-30% in base and 16-25% in non-base hexes.
Main finding #3 (not surprising) is that ABC-divided IDs repair significantly (~2 times) faster compared to whole IDs.
Batch of tests #5: effects of LCU morale on mundane stuff
As per some of Lowpe's questions here tm.asp?m=4265621
80 LCUs of different variety sit (or marched) for several turns with either morale 30 or morale 99 (changed for all of them in scenario editor between tries), all other things being equal.
Main finding Morale does not affect: engineers and avsupport efficiency, marching speed, fatigue/disruption/disablements recovery speed.
Batch of tests #6: effects of HQs
Main finding #1 (not surprising) Presense of HQ/HQc/HQy and the fact of preparation gives significant bonus to LCUs attacking the base. LCU without HQ in range perform at 50% efficiency, that is adjusted AV will be halved on average compared to base AV. HQ in range increases efficiency to 70% even w/o prepping for the base. HQ prepping for the base can increase efficiency up to average 150% (with all units 100 prepped), and up to double in single attacks.
Main finding #2 (mildly surprising) Preparation of attacking LCUs to the same base is necessary to use HQ prepping bonus. Average efficiency increase linearly with LCU prep level (from ~70 to ~150) due to HQ bonus. LCU prep level w/o HQs does not matter much for AV.
If you have any suggestions for other tests, I might be able to run them too. No airgroups though, those are a whole new hassle.
- Attachments
-
- TESTGA34.zip
- (691.66 KiB) Downloaded 33 times
RE: GA statistical musings
Is load cost actually the variable that mattered for the forts? What happens if you take those same IDs and reduce the load cost of the devices within them so that they dropped?
Can you redo #4 and look at what happens if you split IDs for 2-3 days and then recombine them? Do they repair faster?
Can you redo #4 and look at what happens if you split IDs for 2-3 days and then recombine them? Do they repair faster?
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
Good ideas both.ORIGINAL: Lokasenna
Is load cost actually the variable that mattered for the forts? What happens if you take those same IDs and reduce the load cost of the devices within them so that they dropped?
Can you redo #4 and look at what happens if you split IDs for 2-3 days and then recombine them? Do they repair faster?
First: I tried putting most devices at load cost 1, this brought loading costs of standard TOE IDs to 1-7k (forgot some devices but that's ok). Vast majority of standard TOE IDs got to forts 4 in just 12 days. So yes, this is even better example that total loading cost is the thing. Logic might be that you need more time to bury bigger devices into the ground, or you can't dig them down enough at all. Note that it is not about individual devices though, because I had also set some devices (few in the TOE to keep total loading cost low) at loading cost 50, and it did not matter for reaching high forts quick.
Also it does not matter if you LCU is disabled. IDs dug up quickly while still at 100% disabled (from previous tests)
Second: divided IDs repair their disabled devices considerably faster than whole IDs. Like 2 times faster - in 20 days combat (building forts) + 40 days rest in off-base hex regiments repaired 34-46% (when recombined into ID) compared to 13-26% for whole IDs
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
GetAsista,
Some nice studies, thanks.
I have run a couple of small studies re R&D and would like to run some more but I fear that I will die of monotonous reputation if I run the long studies I desire. That is, starting each turn over and over.
Do you have some sort of Batch File to run your turns during your studies to alleviate this?
Some nice studies, thanks.
I have run a couple of small studies re R&D and would like to run some more but I fear that I will die of monotonous reputation if I run the long studies I desire. That is, starting each turn over and over.
Do you have some sort of Batch File to run your turns during your studies to alleviate this?
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
No, just a very small scenario with minimum units, all delays set to zero, and 4 days a turn. The latter does not matter for R&D research as probability does not change while you stay in the same calendar month. So you can read 4 time more repair occurencies each run, essentially having 4 times more data.ORIGINAL: el lobo
Do you have some sort of Batch File to run your turns during your studies to alleviate this?
Also it is handy to look on repairs from Intelligence screen -> aircraft/engine production pool, not from industry screen
Edit: oh, and I changed timeframes by editing aircraft arrival dates and scenario beginning date. Never executed more than 1 starting cycle. Restarted scenario then
RE: GA statistical musings
Can you test one little thing? that is... a Division in terrain split to /A/B/C parts... let it dug in for level 4-5-6 and then recombine. What will be the total fortifications of the whole unit? Is it possible to increase the field fortification levels past the "load coast" limit?
RE: GA statistical musings
ORIGINAL: Barb
Can you test one little thing? that is... a Division in terrain split to /A/B/C parts... let it dug in for level 4-5-6 and then recombine. What will be the total fortifications of the whole unit? Is it possible to increase the field fortification levels past the "load coast" limit?
+1
what happens after you recombine A/B/C in lvl 4 fortification?
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
Sure, tested that already. When all ABC are at Xp forts, recombined division gets X without p (gets p on the next turn but will never build it upwards after that becasue of loading cost limit). I assume ID gets averaged fort level from all ABCs. Did not test what happens when forts are different between ABCs though, assume it is some kind of average too.ORIGINAL: szmike+1ORIGINAL: Barb
Can you test one little thing? that is... a Division in terrain split to /A/B/C parts... let it dug in for level 4-5-6 and then recombine. What will be the total fortifications of the whole unit? Is it possible to increase the field fortification levels past the "load coast" limit?
what happens after you recombine A/B/C in lvl 4 fortification?
This is main way to get your Japanese IDs dig in deeper than 2
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
GetAssistaORIGINAL: GetAssista
If you have any suggestions for other tests, I might be able to run them too. No airgroups though, those are a whole new hassle.
A question I have, and judging by other AARs, others do also.
Re, A/C R&D, what would be the difference time wise in starting R&Ding an a/c from the beginning of the game as opposed to starting R&Ding later in the game.
For example the Frank or Sam. What would the difference be in R&Ding, say fifteen or twenty Locations, from day one, as opposed to starting later say, March of '43. The later scenario would allow you to heavily R&D say the George, and once it comes on-line, switch the Locations to the Frank or Sam and heavily R&D them.
The bottom line is, how many months difference for the later a/c. The engines would also need to be factored-in. (The extra cost will have to decided by the player if it is worth it or not.)
I know running this will take some time to accomplish but it would be nice if you could manage it.
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
Yes, that's a interesting practical question to derive the whole matrix of expected_accelerated_date = M(#factories, frame_time_to_arrival). I just think it is not feasible to do it ingame. E.g. you need to do at least 10 runs from start to finish for each cell to account for lucky/unlucky repairs helping/hindering ongoing research. So it is 10 х 500-900turnsORIGINAL: el lobo
I know running this will take some time to accomplish but it would be nice if you could manage it.
It is preferrrable to do an outside calculation. Given that you have an estimate of probability of unit repairing (and a pretty consistent one I'd say) for each timeframe, you can run simulations e.g. in Excel using this probability
RE: GA statistical musings
Yeah and when you do, you will find it matches in game results. Meaning, you setup 12x30 factories on A7M2 on 7Dec42. Now, the probabilities are that they won't repair until late '43 or early '44 on average. HOWEVER, a couple will be much slower and a couple will be MUCH faster. Those 2 or 3 that repair faster start pulling the date in which causes all of them to repair faster which then get the ac to arrive sooner than the average expected result, if that makes sense. In practical terms, the aggregate solution states that you won't get the A7M2 until late '44 at best. The independent solution shows that you could get it much sooner. Be sure you factor in GA's correct observation that the probability of each factory only seems to update each month, not each day. This is where a good 'roll' can really distort things (ie, it really expands the probability event).
Pax
- Chickenboy
- Posts: 24520
- Joined: Fri Jun 28, 2002 11:30 pm
- Location: San Antonio, TX
RE: GA statistical musings
Here I expected some musings or observations regarding the state of Georgia. All manner of options availed themselves, some of which we would require our most prominent Georgian (Canoerebel) to answer for. [;)]
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
GetAssista and/or other mathematicians.ORIGINAL: GetAssista
P=min[1;30/(months*30-8)]
I would like to run a study using your R&D (tests #1) formula in a spreed-sheet, however I do not fully understand it, in particular, the "min." How would I put that into a formula?
Thanks.
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
Also, why not 22 instead of 30-8? Parentheses?
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
Also I have decided to run this with JavaScript, if that matters.
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
First a note on calculating "months". Assume your airframe arrives at jan-44.ORIGINAL: el loboGetAssista and/or other mathematicians.ORIGINAL: GetAssista
P=min[1;30/(months*30-8)]
I would like to run a study using your R&D (tests #1) formula in a spreed-sheet, however I do not fully understand it, in particular, the "min." How would I put that into a formula?
Thanks.
If you are anywhere in the month of:
Dec-43 then months=1
Nov-43 then months=2
Oct-43 then months=3
... etc
Now put months into the formula:
Dec-43 then 30/(months*30-8) = 30/(1*30-8) = 30/22 ~ 1.36, but the probability cannot be greater than 1, so substitute this with 1. This is what "min" (minimum) is for
Nov-43 then 30/(months*30-8) = 30/(2*30-8) = 30/52 ~ 0.577
Oct-43 then 30/(months*30-8) = 30/(3*30-8) = 30/82 ~ 0.366
... etc
- el lobo
- Posts: 825
- Joined: Fri Jul 19, 2013 9:06 pm
- Location: Khon Kaen, Thailand (orig: Sacramento, CA)
RE: GA statistical musings
Thanks GA. I don't mean to hijack your thread but you kind-of opened it up with your reference to this thread. Besides, I know you are not opposed to discussions so a quick comment.
The chart below is practically useless because alimentary's simple but elegant formula fails to address two things, randomness and the appearance that repair accelerates as the time of availability nears. This is not to disparage alimentary in any way, he is obviously a very intelligent individual. Several times through-out the thread his "program" is mentioned but I did not find it. Perhaps his program addressed these issues.
The only thing that can be derived with any certainty from the chart below is that the sooner the Japanese player starts his R&D on an air-frame, the better. But we already knew that. Like I said, practically useless.
As mentioned, I looked at GA's numbers but they did not quite answer for what I was looking so I worked with alimentary's formula from the thread above.
The chart below is practically useless because alimentary's simple but elegant formula fails to address two things, randomness and the appearance that repair accelerates as the time of availability nears. This is not to disparage alimentary in any way, he is obviously a very intelligent individual. Several times through-out the thread his "program" is mentioned but I did not find it. Perhaps his program addressed these issues.
The only thing that can be derived with any certainty from the chart below is that the sooner the Japanese player starts his R&D on an air-frame, the better. But we already knew that. Like I said, practically useless.
- Attachments
-
- chartfa..irfrank.jpg (139.08 KiB) Viewed 812 times
El Lobo (J) vs Rio Bravo (A)
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
Rio Bravo's AAR - "The War College"
tm.asp?m=3851786
Gamey is not a game style. It is someone's definition.
RE: GA statistical musings
Good job![&o]
I know that Spidery wrote a formula for estimating plane completion...it generated a range: earliest to latest. I am not sure he posted it anywhere, but I am pretty sure he described it somewhere in his AAR.
A great place for Japan to dig in divisions is some of the off base border hexes with the Soviets. Of course most games will never see any actual combat there, and you might have to relocate the border troops to protect Korea.
There are several other important x3 terrain where you can get a division dug in to level 3 or 4 and should be a priority since you know combat will eventually find them.
I know that Spidery wrote a formula for estimating plane completion...it generated a range: earliest to latest. I am not sure he posted it anywhere, but I am pretty sure he described it somewhere in his AAR.
A great place for Japan to dig in divisions is some of the off base border hexes with the Soviets. Of course most games will never see any actual combat there, and you might have to relocate the border troops to protect Korea.
There are several other important x3 terrain where you can get a division dug in to level 3 or 4 and should be a priority since you know combat will eventually find them.
-
- Posts: 2818
- Joined: Sat Sep 19, 2009 6:13 am
RE: GA statistical musings
I've done my own sim some time ago in Excel, so here it is with updated probability from my research here. You can choose date of airframe original arrival, date of research start, with 1 to 10 R&D factories running (with an option to have some factories starting later), with or w/o engine bonus.ORIGINAL: el lobo
Thanks GA. I don't mean to hijack your thread but you kind-of opened it up with your reference to this thread. Besides, I know you are not opposed to discussions so a quick comment.
Edit colored areas except output one and run RND. The latter is done automatically by editing any cell. Output is in terms of simulated airframe arrival date.
I hope all formulas are self-explanatory
- Attachments
-
- research.zip
- (207.26 KiB) Downloaded 25 times
RE: GA statistical musings
ORIGINAL: GetAssista
I've done my own sim some time ago in Excel, so here it is with updated probability from my research here. You can choose date of airframe original arrival, date of research start, with 1 to 10 R&D factories running (with an option to have some factories starting later), with or w/o engine bonus.ORIGINAL: el lobo
Thanks GA. I don't mean to hijack your thread but you kind-of opened it up with your reference to this thread. Besides, I know you are not opposed to discussions so a quick comment.
Edit colored areas except output one and run RND. The latter is done automatically by editing any cell. Output is in terms of simulated airframe arrival date.
I hope all formulas are self-explanatory
Thanks, how do you mod it for less than 10 factories?