433 Haaland = W 609 (60.9%), D 155 (15.5%), L 236 (23.6%), 2.79-1.63 433 Mbappe = W 512 (51.2%), D 196 (19.6%), L 292 (29.2%), 2.23-1.59 433 Osimhen = W 490 (49.0%), D 178 (17.8%), L 332 (33.2%), 2.24-1.68 433 Lewandowski = W 483 (48.3%), D 182 (18.2%), L 335 (33.5%), 2.09-1.65 433 Kane = W 433 (43.3%), D 200 (20.0%), L 367 (36.7%), 2.05-1.74 433 Camavinga (DM) = W 397 (39.7%), D 245 (24.5%), L 358 (35.8%), 1.61-1.48 433 Vlahovic = W 355 (35.5%), D 203 (20.3%), L 442 (44.2%), 1.70-1.83 433 Messi = W 353 (35.3%), D 223 (22.3%), L 424 (42.4%), 1.76-1.82 433 Furuhashi = W 324 (32.4%), D 205 (20.5%), L 471 (47.1%), 1.47-1.72 433 Daizen Maeda = W 294 (29.4%), D 226 (22.6%), L 480 (48.0%), 1.34-1.70 433 Andrian Kraev (DM) = W 294 (29.4%), D 198 (19.8%), L 508 (50.8%), 1.43-1.85 433 Paulo Dybala = W 268 (26.8%), D 200 (20.0%), L 532 (53.2%), 1.42-2.01 433 Control Meta template team = W 241 (24.1%), D 208 (20.8%), L 551 (55.1%), 1.24-1.88
I forgot to mention previously that in my test I'm using the following settings:
Tested players are in home team
Meta team - This is to control things better and provide a point of difference to blackbird's results which used normal teams. I plan on doing player rank testing with normal teams as well.
433 V7 variant tactic - Initially I accidentally used it instead of Knap, but then I figured may as well use it to see if there's any differences, since all of my previous manual test data used the 424 Knap tactic which can be compared to. 424 will be tested later as well.
Attendance 1 Home/1 Away/1 Season Ticket Holders - So that homeground advantage doesn't cloud things, but this may mean the results may be invalid in regards to pressure and important matches
Condition ~96-98%, Match Sharpness 98.5% (much less on bench), no fatigue, ~60% morale - I wanted make it realistic, particularly as at least dirtiness seems to interact with fitness levels, and fitness levels definitely has a big impact on outcomes. I also haven't bothered to change a bit of minor variation among players in the condition or whatnot. This is a little bad for pinning down the exact effect of attributes, but otherwise I feel it might be slightly beneficial in ways. For example, what if the amount of variance of player condition % within a team affects the outcomes. The safest bet seems to me to go with what is realistic, so that it captures any unknown unknowns.
I've chosen to play against Tottenham, with their team having decent fitness levels. If you choose a team like Liverpool, it's likely you're favoring attributes that will only work against a minority of teams that will beat you (say ~20%). There is no fully representative team, but I figured picking a more middling team is more likely to represent the bulk (say ~80%). Better to win most of 80% and lose against 20%, than to win most of 20% and lose against 80%. I feel like Tottenham is a good inbetween choice.
In the player ranking results we see that it lines up fairly well with the previous results, but there are a few surprizes.
Maeda and Furuhashi drop down quite a lot. I suspect this is actually because of the test conditions. I will see later how Maeda does with conditions changed to normal team, weaker opposition, Knap tactic. It could also be that dirtiness say is having less impact because its 1 match being repeated rather than a full unfolding season (i.e. yellow cards dont matter anymore).
Another thing that I've marked is that Camavinga (DM) performed as well as expected, but the average scoreline is noticeably unique. It seems to reveal that STs mainly score more, while DMs mainly concede less, even though they share the same meaningful attributes and similar win rate effects from those attributes.
I'd be willing to believe that Osimhen edges out Lewandowski, its not unthinkable, but Maeda being that low doesn't seem right to me. I suspect we're going to see things change with different test conditions, but we'll see.
Attribute testing is like all these health studies that tell you 2 eggs/week = +14% death risk
Calibrating weights to real player rankings is like eating the Mediterranean diet because people who eat it live longer and are generally healthier.
Sometimes the two contradict each other, but they're probably largely consistent with each other if only you were aware of the nuance.
And what we can do is use one to tell us about the other. Perhaps one is more valid than the other, but nonetheless both kinds of information remain important.
Meanwhile LightingFlik just came up with the AI that can cure cancer, but we're using it to verify the basics first, like what foods are best to eat and just how healthy are eggs exactly.
1) Acceleration, Pace and Dribbling are the most valuable attributes by FAR for Outfield and if you look for them for your Outfields then you're safe.
2) Reflexes, Agility and Acceleration are the most valuable attributes by FAR for GKs and if you look for them for your GK's then you're safe.
And you're getting a very similar pictures here.
P.S. You should also tested Agility for GK, it did well in FM-Arena's test. Expand *cracks knuckles*
I kind of had a feeling people would interpret it this way - I had in mind that some people are no doubt going to unwittingly take this as a set of weights to plug into Genie Scout, which is basically the same error.
What I found since those 2 years ago is that real player rankings do not line up with the attribute test data.
This is why on the first page my attribute template has dribbling at '13' and work rate at '11', while in my FMST26 weights just below it, dribbling is 14.3 and work rate is 66.5. We have evolved beyond individual attribute testing as the source of weights, or at least I have. The method I came up with was to change the attribute weights around until it fit as closely as possible the real player results - and I've tested the resulting weights and found them to be superior.
Now someone might chime in and say 'well I'm not interested in weights, I just wanted to know what going from '8' pace to '20' pace results in'. To that I'd say good for you, but in regards to what I'm doing, I'm going beyond that.
Now, returning to the present.. Why am I doing attribute tests now then?
The plan is this:
1) Utilize LightningFlik's program to more accurately identify attribute result curves, and previously unknown attribute combinations
2) Use the program to also rapidly do mass testing of individual players for the real player ranking list
3) Use what turns up out of it all to make better informed guesses for the selection of weights to try and match real player rankings
Attribute testing isn't the main benefit of LightingFlik's tool IMO, it's stuff like what's he posted finding about how the game works, and the ability to do mass tests of player swaps with ease, and perhaps a lot more.
The ability to do attribute testing 20+ times faster is nice, but not exactly a game changer. You could already do it before, and it was done - by FM Arena, Harvest Green, myself, etc. - it's just about how much time are you willing to sink into it.
That said, the ability to do 1000 samples in 3 minutes or so does enable some previously unrealistic things, like getting accurate read on the GK attributes. Even my own manual findings on GK have not been that satisfactory.
I did do agility, I forgot to add it in. I'll add it in now.
Edit: Actually I see why I left agility out. I've got mixed results on it. What we can glean from that at least is that agility probably doesn't matter that much, but I want to do some more testing on it to provide further clarification.
Reflexes = 19.88 Acceleration = 12.10 Pace = 12.10 Command of Area = 8.53 Determination = 3.97 One On Ones = 2.31 Aerial Reach = 2.19 First Touch = 2.19 Rushing Out (Tendency) = 1.74 Punching (Tendency) = 1.70 Concentration = 1.46 Technique = 0.41 Handling = 0.20 Throwing = -0.24 Communication = -0.45 Eccentricity = -2.55 Jumping Reach = -3.63 Strength = -4.11 Kicking = -4.41 Stamina = -6.70
Notes:
The attributes are represented as relative to each other. +1 reflexes on the GK is equivalent to +0.2 pace on ALL outfield players.
I count +1 for win and +0.333 for draw (before conversion), that's all I'm judging by.
The negative GK attributes are not mistakes from one-off outliers. For instance for Strength here's the tests and trend:
Strength 1 (1000 samples) = W 262 (26.2%), D 209 (20.9%), L 529 (52.9%), 1.29-1.85 Strength 10 (1000 samples) = W 257 (25.7%), D 203 (20.3%), L 540 (54.0%), 1.27-1.84 Strength 12 (1000 samples) = W 252 (25.2%), D 212 (21.2%), L 536 (53.6%), 1.27-1.86 Strength 16 (1000 samples) = W 244 (24.4%), D 204 (20.4%), L 552 (55.2%), 1.23-1.83
I'm being selective about what I test. Some I just do '1' and '18'. If the result is unusual, like Strength, or crucial, like Reflexes, I do more intervals. I haven't done this for 1 or 2 yet though such as GK stamina.
Because attributes are non-linear, I've decided to represent the results here as more typical/realistic values rather than simply '1' vs. '20'. A few examples to illustrate what some of the above figures are derived from:
I've found a few (relatively minor but notable) differences to blackbird's results so far. For instance aggression does matter, but mostly only going from 1 > 6 (blackbird only did 4 > 8, and higher). I did not see a significant difference for Strength going from 10 > 20, but 6 > 10 is significant. I'll be posting the extended results later that bears this out.
It could partly be that blackbird used a tactic that affects the outcomes. I've used a variant of the Mars Attacks V7 tactic (433 formation) for my testing. I've previously done all my testing with the Knap tactic, so when I realized I was using the V7 tactic, I figured I'll just continue with it anyway since that will shed light at the same time on whether tactics/formations affect attribute outcomes or not. And it doesn't seem that they do.. so long as they are an intense gegenpress style tactic at least I guess.
Later on my results should be able to mostly complement blackbird's results to give a more complete picture (i.e. values for 1, 4, 6, 8, 10, etc.)
LightningFlik said: If there's something I can do to make it easier let me know. I'm currently thinking of making it possible to insert players in to a team (I eanto see how far a squad of Haalands gets) and edit players in the UI. Expand It's ok, I think mainly I just need to take a break and then understand properly how to use it. I had that crash issue, but I can mostly avoid this so it's not really that.
If no one else was doing the tests, I'd probably push myself to do it, but since you & others are getting results (and more sophisticated than what I can currently do), no point putting myself through tedium. The benefits of this are so good that I will get back into it after a brief break though.
What made me call it a day is I edited the attributes in the JSON as you suggested, creating a changed JSON and a control JSON which did get different results, but then when I edited the control JSON it kept producing identical results when it shouldn't. And the problem is, I don't know if this is you're using the wrong methodology (i.e. maybe you don't realize its already locked in certain attributes at this point), or if I'm doing something wrong. I suspect it's the latter, but I can't tell. So I really need to develop my understanding first.
It did occur to me writing my last post that the program can also be applied to the real player alignment process too, which will speed it up in the same way it sped up attribute testing.
I've done an outfield squad of Haalands btw, with manual testing. Scores around +150 g/d in the Premier League with Knap tactic. For reference, the max possible is somewhere around +450 g/d.
Someone asked me about the GS weights, and I thought I'll post here as well what I said about it.
LightningFlik's program is giving us more precise data to use, but so far it doesn't look like much needs to change.
Any new changes probably still need to go through the same process of lining up performances of real players correctly and deriving weights from that. Having more precise attribute data is going to improve this, because it will narrow down the possible combinations that explain things, but this manual adjustment to line up real players still has to be done using the data.
A change to the GK weights might happen quicker though, as I think it's likely to have a few errors that will be uncovered through this new attribute testing method.
---
Is anyone doing the GK attributes?
I'm still having some troubles with getting the testing working properly or easily, and I'm quite burnt out by fiddling around with it. So I probably won't be contributing much immediately, especially with others having a better go of it than me.
This test shows that some attributes are indeed non-linear Expand This all just about lines up with what I found through manual testing, so there's not much to change in my ideal template I think. I didn't have this level of precision though of course.
There's a few small surprises from the new data so far:
Aggression (apparently useless) Strength (at 20) Finishing (at 16-20) Dribbling (at 4-8) Dirtiness (seems to have an unusual mechanic, I wouldn't go by the results here for this one) Vision (at 4) Determination (at 8)
I kind of already knew that strength and finishing had benefits at higher levels, I surmised that it's not worth the extra CA, but I thought it was more linear than what the results show here. I still don't think they're worth getting though. Perhaps players who happen to have very high finishing are more worthwhile though.
Determination seems to work not too bad at lower levels, but there's also the effect on training to consider, so I think higher is still better.
If we surmise pace as the difference between 12 and 20 pace, it's 1.35/8 = 0.16875 per 1 pace
So
20 strength = 2.19 pace 20 finishing = 2.32 pace 20 longshots = 1.48 pace
20 agility = 2.30 pace 20 balance = 2.13 pace
Strength, agility and balance are very similar, but agility has a much more severe impact when 8 or lower. Balance is inbetween, which is a little surprising, because my testing found balance to be a bit more important than agility.
Crossing seems to carry marginal improvements up to ~16. Crossing 8 > 16 is equivalent to +1.70 pace. For comparison, Strength 12 > 20 is +1.37 pace, so that's a considerable benefit. Other attributes are more important of course. Consider 8 > 12 composure for instance, which is +1.29 pace for half the attribute points (so +2.58 pace for composure vs. +1.70 pace for crossing).
I did a few more tests yesterday, which I'll post since they include temperament and professionalism which blackbird hasn't included in his test:
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Teamwork 14 Meta vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 52 (17.33%), Loss 31 (10.33%), avg. 2.47-0.88
Temperament 20 Meta vs Man Utd (Home) Fully healed 300 sample = Win 222 (74.00%, Draw 43 (14.33%), Loss 35 (11.67%), avg. 2.44-0.82
Professionalism 8 Meta vs Man Utd (Home) Fully healed 300 sample = Win 220 (73.33%), Draw 51 (17.00%), Loss 29 (9.67%), avg. 2.50-0.89
Natural Fitness 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 213 (71.00%), Draw 56 (18.67%), Loss 31 (10.33%), avg. 2.36-0.82
Determination 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 192 (64.00%), Draw 60 (20.00%), Loss 48 (16.00%), avg. 2.39-1.00
98con/97fit/75morale Meta vs Man Utd (Home) Fully healed 300 sample = Win 186 (62.00%), Draw 68 (22.67%), Loss 46 (15.33%), avg. 2.16-0.92
So it seems that temperament 20 might not be worse than temperament 10 as Orion found. It doesn't seem to have any benefit over 10 either though.
Professionalism surprisingly seems to have little downside going from 13 > 8. Of course it matters for training and other things.
The natural fitness result seems to align with blackbird's result that it makes no difference. But it is nonetheless very important because it affects match sharpness loss rate outside of the matches, which is very impactful on performance.
Even a slight change in fitness and/or morale has a significant impact on results. So trying to get that extra ~2% condition or little boost to morale prior to a match really does matter. We see here -2% condition & -25% morale is worse than 14 > 1 Determination for the entire outfield.
LightningFlik said: 100% home fans: home 588 / draw 181 / away 231, avg 3.38-2.36 100% away fans: home 234 / draw 173 / away 593, avg 2.18-3.19 50%/50% fans: home 399 / draw 191 / away 410, avg 2.72-2.73 No fans: home 407 / draw 199 / away 394, avg 2.70-2.74 75/25 fans: home 389 / draw 195 / away 416, avg 2.72-2.79 75,000/25,000 fans: home 512 / draw 202 / away 286, avg 3.08-2.46
(These last two tests were to see if the absolute value of fans made the positive impact or the relative strength. Given the absolute number of fans makes a difference then there are performance gains in upgrading your stadium as much as possible, since every fan counts.) Expand Bruh what
Fan count affecting results was not something I was expecting
I guess this has implications for team selection when playing
LightningFlik said: Are you changing attributes in the game or the fixture object? If you edit the fixture JSON, changing the attributes should be all you need unless I'm very much mistaken.
(And, if I am, I can bake in the recalculation.) Expand I'm using a realtime editor to change the attributes in the game before exporting to JSON
I haven't taken the time yet to examine how your program works. So I don't know what you mean by 'fixture object'. Are you saying we can just edit the attribute data in the JSON file after export?
Let me know the process, because having to change the CA for each player is a real pain
LightningFlik said: Why do you need to adjust CA? I don't think it affects a player's performance. Expand If I change an attribute that has CA weight, I have to change the player's CA to the new Recommended CA, otherwise it messes with the results.
I had my fingers crossed initially that the game would automatically just recalculate the CA for the player, but its not the case. It reduces every other attribute a little to make up for it.
When I added the correct CA, I got results that lined up with expectation (i.e. teamwork being marginally beneficial, rather than 8 > 14 teamwork producing significantly worse results)
Jolt said: Ultimately, one team dominating the other is not a measurable metric (what threshold - and what metrics define the trigger - qualifies as "dominating"? noone knows, thus it is not useful), so unless the reverse engineering allows insight into it, it is not a very useful scenario. Expand Dominating means winning almost all games with ease, which we measure with the points tally.
There is a high correlation with fitness and physical attributes. We don't know what it's specifically composed of, yes, but I wouldn't go so far as to say it's not measurable.
Jolt said: What @LightningFlik mentioned was that certain groups of attributes may be made more effective by an external attribute threshold (Concentration >11). Expand I think you've made an error of interpretation here. I think LightningFlik is saying that concentration >11 was necessary for certain miniscule match focus boosts, not for groups of attributes to kick in or be more effective.
It's possible similar things are going on with attributes, we don't have anything to suggest that yet.
One thing that suggests its not going to be like that is that I scaled down my template for lower leagues, and it seemed to work just as good. So I think it's about attributes relative to the opposition, rather than set thresholds for attributes to take effect.
Jolt said: Of course, this is an utterly gargantuan amount of work. All visible attributes, plus the four hidden ones we know have an effect on performance (Consistency, Important Matches [this one would require editing match reputation or editing rivalry between the teams], Pressure, Dirtiness), and we're talking about close to 10,000 3-attribute combinations.
Even testing just 3 variations per attribute (7, 12, 17), and running 10 games per variation to get a sample size, and we're talking about hundreds of thousands of games.
So it is understandable that noone followed up on it. Expand I think it's viable to do in a certain way
We can get rid of some attributes such as long throws, penalty taking, etc.
Then we can also discard some combinations such as vision + natural fitness + technique. And order of the combination doesn't matter either.
I did about 11,000 samples last night. Not sure on the exact numbers, but this means our processing capacity just increased at least 20-fold. Maybe applying AI to it can strip away the remaining problem of tedium.
I can't remember what I got this from, but I think combos can supposedly be up to 5 attributes.
What's holding me back right now is that it's super tedious to adjust CA in line with attribute changes in the realtime editor. Maybe the official realtime editor automatically does this, might see if I can use it later.
- In possession: Off the Ball is rolled twice every 15 seconds (instead of just once) when your team has the ball and the higher of the two values is used. - In Transition: Teamwork is re-rolled after possession change and the higher of the two values is kept. - Out of Possession: Marking is rolled twice while the opponent has the ball and the higher of the two values is used. - Attacking corners: Anticipation is rolled twice when you have a corner and the higher result is used - Defending corners: same as above except when the opponent gets a corner - Attacking free kicks: same as above - Defending free kicks: same as above - Throw-ins: Re-rolls anticipation or gives a +1 bonus to the original value if it was already better - Penalties: a flat +1 bonus when taking a penalty
And by the way, I think that to get these bonuses, a player's current Concentration roll has to be >11, which means players with a low concentration attribute never see these buffs. Expand
So it sounds like match focus is kind of like a consistency modifier for those particular attributes you mention. Sounds like 'transition phrase' would be the one to pick, since teamwork actually does a little something.
I think I can shed some possible light on this. My own testing has shown that there are certain combinations of technical attributes (i.e. crossing + long shots + technique) that produce a significant difference, but this only came through when the team was dominating.
I think what this means is that when Man City play Barrow in the FA Cup, they're going to go above and beyond what Burton would do to Barrow, but only so long as their physicals mean they can keep it up - and in this case, they should have enough physical superiority to keep it up all game, whereas Burton might have to call it a day at 5-0 60 minutes in.
But if its say Man City vs. Luton, Man City should start with enough physical advantage to execute all that technical finesse for a while, but then I think the idea is that it ends up becoming a slogfest for much of the match where whoever makes the least serious mistakes (i.e. low dirtiness, high concentration) wins the battle.
IMO I think it's deliberate but probably unintentionally unbalanced, because it's common sense that if a player is worn out, he's not going to be pulling off clever technical skills.
With these combinations not showing up in tests, it would likely be because these tests are controlled where all players about equal in fitness all through the match due to near identical attributes, healing, etc.
I used the harness to brute force test almost all possible settings for my 'Variant B' of the Mars Attacks V7 tactic with my meta team.
I cumulatively improved it from W 132, D 36, L 32, avg. 2.02-0.94 to W 142, D 39, L 19, avg. 2.21-0.79.
For comparison, the Knap 424 scored W 135.2, D 39.2, L 25.6, avg. 2.36-0.86.
I then tested it with Rushall, finished 3rd. I thought a bit disappointing, but not too bad.
I tried Luton next. I got 1st with an unselected team straight off the bat. I thought bloody hell, this is it. To get 1st with Luton with a selected team is very rare. To get it with an unselected team is unprecedented. I think I was even a few points clear of 2nd.
But then I got 6th and 10th. And with Man City, the results were 1st but only 93pts and 100pts.
But today I see that I accidentally used what scored W 135, D 39, L 26, avg. 2.35-0.89 instead.
After fixing it up, I've still only got 7th as unselected Luton and 4th as selected Luton as the two initial samples.
So perhaps tactics are going to be somewhat safe from this brute-forcing method. This and the dirtiness results suggests to me that the non-linear nature of attributes and their interaction with external factors such % condition, means that we can't simply universalize a tactic that works perfectly for one team. Each team would have a slightly different optimum, and the method seemingly can't really tell you what tactic works best with most teams overall.
But I guess what I can try next is to brute force Luton rather than the meta team, and see what that produces.
The results are suggesting that it is in fact the case that the negative of dirtiness only kicks in at lower fitness levels:
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Dirtiness 12 Meta vs Man Utd (Home) Fully healed 300 sample = Win 209 (69.67%), Draw 54 (18.00%), Loss 37 (12.33%), avg. 2.39-0.94
Dirtiness 12 Meta (1 day before) vs Man Utd (Home) Fully healed 300 sample = Win 197, Draw 67, Loss 36, avg. 2.33-0.91
We see above that Dirtiness 8 > 12 seems to make no difference when the squad is set to 100% fitness.
Longshots Technique Passing Crossing 16 vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 98, Loss 110, avg. 1.35-1.49
Dirtiness 12 Longshots Technique Passing Crossing 16 vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 79, Loss 129, avg. 1.35-1.62
By contrast, when fitness is left to standard levels - around ~90% condition ~100% match fitness on average in this case - there does seem to be a significant difference.
This will require further testing to clarify. Other attributes aren't working this way, and are producing results in line with expectations.
I believe that the new method I'm using has fixed up my previous teamwork & technicals results, which should be disregarded as erroneous. I think it's mainly that I didn't adjust CA upwards in the realtime editor to match the increases, which I've done accurately now. I'll probably edit the previous post later to remove the old results to avoid confusion.
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Teamwork 14 Meta vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 52 (17.33%), Loss 31 (10.33%), avg. 2.47-0.88
So 8 > 14 Teamwork is now seemingly marginally beneficial. This is what I would have expected.
Pressure 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 146 (48.67%), Draw 73 (24.33%), Loss 81 (27.00%), avg. 1.60-1.00
Thought I'd try a very obvious one just to make sure things are on the right track and provide a yardstick to compare these lesser attributes to. Big drop from Pressure 13 > 1 as expected.
Longshots Technique Passing Crossing 16 Meta vs Man Utd (Home) Fully healed 300 sample = Win 236 (78.67%), Draw 48 (16.00%), Loss 16 (5.33%), avg. 2.88-0.84
Control Meta (Knap 424 tactic) vs Wolves (Away) Fully healed 300 sample = Win 256 (85.33%, Draw 31 (10.33%), Loss 13 (4.33%), avg. 2.75-0.83
Longshots Technique Passing Crossing 16 Meta vs Wolves (Away) Fully healed 300 sample = Win 274 (91.33%), Draw 21 (7.00%), Loss 5 (1.67%), avg. 3.09-0.82
Longshots Technique Passing Crossing 16 Meta vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 98, Loss 110, avg. 1.35-1.49
These attributes performed a quite a bit better than I expected. I did find in my testing that this particular combination yields significantly higher goal difference at very high performance levels, but this is showing significant benefits even at a lower performance level - albeit at a cost of +32 total attribute points and a good deal of CA to match.
Control Meta (Knap EF 424 IF HP V2 P101 AC) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
V7 Variant B vs Man Utd (Home) Fully healed 1000 sample = Win 656 (65.6%), Draw 221 (22.1%), Loss 123 (12.3%), avg. 2.14-0.86
424 Mountain King vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 208 (69.33%), Draw 63 (21.00%), Loss 29 (9.67%), avg. 2.36-0.78
424 TON 424 HP HUB MC P105 EFC vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 221 (73.67%), Draw 50 (16.67%), Loss 29 (9.67%), avg. 2.15-0.75
424 'Park the Bus' (default tactic) vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 50 (16.67%), Draw 78 (26.00%), Loss 172 (57.33%), avg. 1.37-0.66
I think an issue with comparing tactics is that the players you are fielding matter, and in real play you're more likely to get good players in certain positions rather than others.
So we can say, this tactic does best in a controlled test, and that tactic does best in a default plug-and-play with Luton, but it would be difficult to find conclusively that X tactic is 2% better than Y tactic. But you could demonstrate it's 2% better in 5 or 6 scenarios say.
Meta team Teamwork 14 (from 8) Fully healed 200 sample = Win 107 (53.23%), Draw 50 (24.88%), Loss 44 (21.89%), 1.93-1.00
Meta team Determination 8 (from 14) Fully healed 200 sample = Win 112 (55.72%), Draw 44 (21.89%), Loss 45 (22.39%), 1.93-0.96
Control reference: 63.0% Win, 24.2% Draw, 13.0% Loss
The Teamwork result could be wrong, as with other CA-changing attributes I've done. I will redo all these with a better method and proper CA edits. But from this result, it suggests Teamwork is a negative attribute.
The non-CA attribute results should all be about right though, which includes Determination.
A little surprisingly lower determination isn't too bad, at least in this range. 14 > 8 Determination is equivalent to 8 > 12 dirtiness.
ZaxanR said: Question - is it possible to mod the way that the game highlights player attributes to make them align to the recommended attributes in the opening post? Expand Since no one has answered you, I'll say no way that I know of, but perhaps there's a way of editing the skin to do it.
Kma said: Helllo please @GeorgeFloydOverdosed is there some sort of hidden attribute or stat that will determine players to become staff or not or it's totaly random? have you test that?
and what that numbers meaning in the pre game editor? 95% of players are like that? Expand I haven't tested this. There's probably a few factors involved. That section in the editor is I think just how willing they'd be to accept the particular role if you make them an offer, but there would have to be other stuff influencing them to become staff member instead of retire from game entirely, and so forth.
LightningFlik said: Pretty sure these relationships have no effect in matches by the way.
I've found the data that represents them in memory: it tracks games and minutes played together (plus two other values I've not deciphered yet). But I can't see these in the fixture data, meaning they don't factor in to on-field performance. Expand Yeah I mentioned this in my tactic test findings, so I can confirm this would be true
LightningFlik said: Good question. I ran my experiment using on tiniest database, EPL only, managing default Aston Villa, using Knap's 424 tactic, playing a friendly at the earliest opportunity against Bristol City. (The exact lineup I used probably doesn't matter, but suffice to say I let the assistant pick everything.)
You could recreate those conditions if you wanted to double check my result.
If you match my result then the discrepancy might come from a quirk of your meta-players' attributes. If you don't match my result then I can re-run my experiment later. Maybe I had incorrectly mapped the anticipation attribute in my test or maybe I was still simulating games incorrectly (for a time the player's manager would make no subs).
Also you could pastebin me your fixture JSON and I could re-run my experiment on that and see if I get the same result as you. Expand I thought the tactic could be part of it, but apparently not, as we've used the same tactic.
If it was pace/acc, then it being a quirk of the meta player's attributes would make sense - as pace/acc is on a cliff edge for the meta team, but anticipation shouldn't be like that as it's fairly linear I think. But I did check Aston Villa's players because maybe they were all going from '16' > '11', where '11' wouldn't be much worse probably. But I checked a few and they were mostly ~14, so it would be '14' > '9', which is similar to what I did.
So I reckon it's some methodology issue. Main thing is it doesn't seem to be an issue with the program now.
So we see here that Aggression doesn't seem to do anything. If anything, a very slight gain from lower aggression.
The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier.
It's also surprising to me how strong the RNG is. Even comparing 100 match samples to 500, there's a 10% difference. Though perhaps I need to adjust the way I'm testing for a more controlled result.
LightningFlik said: It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game? Expand I was going to say maybe Leadership is keeping morale of other lesser players high, but then I realized that that's one thing that can probably be ruled out, as I'm giving them all identical attributes and leadership.
And the same could be said for the 'leader' boosting other's determination, etc. to his level.
Using LightningFlik's program, I have done some testing of 'dirtiness'. I think I'm doing it right.
Vs. Nottm Forest (Man City away) (meta team, but with 16 leadership - forgot to change that):
Meta team Dirtiness 12 = Win 45, Draw 26, Loss 30, avg. 1.55-1.18 Meta team Dirtiness 10 = Win 38, Draw 28, Loss 35, avg. 1.45-1.27 Meta team Dirtiness 9 = Win 43, Draw 33, Loss 25, avg. 1.51-1.02 Meta team Dirtiness 8 = Win 50, Draw 34, Loss 17, avg. 1.82-1.13 Meta team Dirtiness 7 = Win 45, Draw 27, Loss 29, avg. 1.64-1.27 Meta team Dirtiness 5 = Win 41, Draw 28, Loss 32, avg. 1.48-1.17 Meta team Dirtiness 1 = Win 44, Draw 24, Loss 33, avg. 1.44-1.07
All of those are 100 samples each and use standard fitness, morale, etc. values.
Meta team Dirtiness 8 Fully healed 100 sample = Win 69, Draw 15, Loss 17, avg. 2.20-0.76
Meta team Dirtiness 1 Fully healed 500 sample = Win 321, Draw 107, Loss 73, avg. 2.06-0.89 Meta team Dirtiness 8 Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89 Meta team Dirtiness 12 Fully healed 500 sample = Win 276, Draw 119, Loss 106, avg. 1.97-1.06
As percentages:
Dirtiness 1 = 64.2% Win, 21.4% Draw, 14.6% Loss Dirtiness 8 = 63.0% Win, 24.2% Draw, 13.0% Loss Dirtiness 12 = 55.2% win, 23.8% Draw, 21.2% Loss
In the 100 sample tests, 8 dirtiness performs best. In the 500 sample one, 1 dirtiness just about equals 8 dirtiness. 12 dirtiness is clearly worse than 8 dirtiness in both.
I can't recall what testing I did exactly which led me to believe this, but I think I have found at some point that 1 dirtiness did worse than 8 dirtiness, so I did figure we have a kind of bathtub curve going on with dirtiness. But on the other hand, this leaves players like Maeda inexplicable. So I ended up just setting it aside but having it in the back of my mind, and it's one reason I wanted to test dirtiness first.
And so I'm not surprised by these results showing 8 is optimal, though I would have expected it to be 6 rather than 8.
What's confusing though is that in the larger sample with perfect fitness/morale/etc 8 does the same as 1. Maybe the effect of dirtiness changes according to fitness condition of players involved say. Or maybe 8 really is the same as 1 and there was too much randomness in the fewer samples.
Thinking about it.. Maeda has really high pace/acc, so maybe he benefits from avoiding doing dirty things, because then he takes advantage of his high pace/acc instead. But if a player has low pace/acc, or is tired, he's better off doing something dirty to get something out of the game. Maybe that's what going on.
In any case, the data does seem to support the fact at least that lower dirtiness = generally better results, with distinguishable significance. We'll see how it compares to another attribute. I think I'll test leadership next.
Edit:
Meta team Leadership 8 Fully healed 500 sample = Win 307, 105, Loss 89, avg. 2.01-1.00
Compare to:
Meta team Dirtiness 8 (Leadership 16) Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
So +8 leadership seems to have slightly better results. About half the difference of dirtiness 8 > 12. So 2 dirtiness = 8 leadership? Will test aggression next to compare.
Before this I did test Leadership again and got results of 3rd, 4th, 5th with Leadership @ 16 instead of 8. So it seems like leadership does have some effect. Not sure how I missed it, I probably only tested it once, got a 5th (which means no position change) and moved on. My guess now is it would have a weight of ~10 maybe.
LightningFlik said: Right, have at it: v1.0.0. Let me know how you get on and I'll try to iron out any bugs as they're reported. After a few days, I'll shout louder about this but you can be my guinea pigs for now. Expand Needed to use ././ to get it working
But then also:
# ././fm24ctl get-next-fixture next-fixture.bin fm24_harness_p10.dll not found; injecting fm24_harness.dll (it must be a protocol 10 build) injected into fm.exe (pid 22960); see harness.log for the build check the harness did not start in fm.exe (wrong game build?); see harness.log
Log:
10:55:26.461 loaded into fm.exe at 0000000140000000 10:55:26.461 refusing: fm.exe timestamp 1744804949, expected 1744804480 (different game build)
433 Haaland = W 609 (60.9%), D 155 (15.5%), L 236 (23.6%), 2.79-1.63
433 Mbappe = W 512 (51.2%), D 196 (19.6%), L 292 (29.2%), 2.23-1.59
433 Osimhen = W 490 (49.0%), D 178 (17.8%), L 332 (33.2%), 2.24-1.68
433 Lewandowski = W 483 (48.3%), D 182 (18.2%), L 335 (33.5%), 2.09-1.65
433 Kane = W 433 (43.3%), D 200 (20.0%), L 367 (36.7%), 2.05-1.74
433 Camavinga (DM) = W 397 (39.7%), D 245 (24.5%), L 358 (35.8%), 1.61-1.48
433 Vlahovic = W 355 (35.5%), D 203 (20.3%), L 442 (44.2%), 1.70-1.83
433 Messi = W 353 (35.3%), D 223 (22.3%), L 424 (42.4%), 1.76-1.82
433 Furuhashi = W 324 (32.4%), D 205 (20.5%), L 471 (47.1%), 1.47-1.72
433 Daizen Maeda = W 294 (29.4%), D 226 (22.6%), L 480 (48.0%), 1.34-1.70
433 Andrian Kraev (DM) = W 294 (29.4%), D 198 (19.8%), L 508 (50.8%), 1.43-1.85
433 Paulo Dybala = W 268 (26.8%), D 200 (20.0%), L 532 (53.2%), 1.42-2.01
433 Control Meta template team = W 241 (24.1%), D 208 (20.8%), L 551 (55.1%), 1.24-1.88
I forgot to mention previously that in my test I'm using the following settings:
Tested players are in home team
Meta team - This is to control things better and provide a point of difference to blackbird's results which used normal teams. I plan on doing player rank testing with normal teams as well.
433 V7 variant tactic - Initially I accidentally used it instead of Knap, but then I figured may as well use it to see if there's any differences, since all of my previous manual test data used the 424 Knap tactic which can be compared to. 424 will be tested later as well.
Attendance 1 Home/1 Away/1 Season Ticket Holders - So that homeground advantage doesn't cloud things, but this may mean the results may be invalid in regards to pressure and important matches
Condition ~96-98%, Match Sharpness 98.5% (much less on bench), no fatigue, ~60% morale - I wanted make it realistic, particularly as at least dirtiness seems to interact with fitness levels, and fitness levels definitely has a big impact on outcomes. I also haven't bothered to change a bit of minor variation among players in the condition or whatnot. This is a little bad for pinning down the exact effect of attributes, but otherwise I feel it might be slightly beneficial in ways. For example, what if the amount of variance of player condition % within a team affects the outcomes. The safest bet seems to me to go with what is realistic, so that it captures any unknown unknowns.
I've chosen to play against Tottenham, with their team having decent fitness levels. If you choose a team like Liverpool, it's likely you're favoring attributes that will only work against a minority of teams that will beat you (say ~20%). There is no fully representative team, but I figured picking a more middling team is more likely to represent the bulk (say ~80%). Better to win most of 80% and lose against 20%, than to win most of 20% and lose against 80%. I feel like Tottenham is a good inbetween choice.
In the player ranking results we see that it lines up fairly well with the previous results, but there are a few surprizes.
Maeda and Furuhashi drop down quite a lot. I suspect this is actually because of the test conditions. I will see later how Maeda does with conditions changed to normal team, weaker opposition, Knap tactic. It could also be that dirtiness say is having less impact because its 1 match being repeated rather than a full unfolding season (i.e. yellow cards dont matter anymore).
Another thing that I've marked is that Camavinga (DM) performed as well as expected, but the average scoreline is noticeably unique. It seems to reveal that STs mainly score more, while DMs mainly concede less, even though they share the same meaningful attributes and similar win rate effects from those attributes.
I'd be willing to believe that Osimhen edges out Lewandowski, its not unthinkable, but Maeda being that low doesn't seem right to me. I suspect we're going to see things change with different test conditions, but we'll see.
Calibrating weights to real player rankings is like eating the Mediterranean diet because people who eat it live longer and are generally healthier.
Sometimes the two contradict each other, but they're probably largely consistent with each other if only you were aware of the nuance.
And what we can do is use one to tell us about the other. Perhaps one is more valid than the other, but nonetheless both kinds of information remain important.
Meanwhile LightingFlik just came up with the AI that can cure cancer, but we're using it to verify the basics first, like what foods are best to eat and just how healthy are eggs exactly.
That's my take on it.
https://fm-arena.com/thread/14009-attribute-testing-football-manager-24/
From FM-Arena's test we learned that:
1) Acceleration, Pace and Dribbling are the most valuable attributes by FAR for Outfield and if you look for them for your Outfields then you're safe.
2) Reflexes, Agility and Acceleration are the most valuable attributes by FAR for GKs and if you look for them for your GK's then you're safe.
And you're getting a very similar pictures here.
P.S. You should also tested Agility for GK, it did well in FM-Arena's test.
*cracks knuckles*
I kind of had a feeling people would interpret it this way - I had in mind that some people are no doubt going to unwittingly take this as a set of weights to plug into Genie Scout, which is basically the same error.
What I found since those 2 years ago is that real player rankings do not line up with the attribute test data.
This is why on the first page my attribute template has dribbling at '13' and work rate at '11', while in my FMST26 weights just below it, dribbling is 14.3 and work rate is 66.5. We have evolved beyond individual attribute testing as the source of weights, or at least I have. The method I came up with was to change the attribute weights around until it fit as closely as possible the real player results - and I've tested the resulting weights and found them to be superior.
Now someone might chime in and say 'well I'm not interested in weights, I just wanted to know what going from '8' pace to '20' pace results in'. To that I'd say good for you, but in regards to what I'm doing, I'm going beyond that.
Now, returning to the present.. Why am I doing attribute tests now then?
The plan is this:
1) Utilize LightningFlik's program to more accurately identify attribute result curves, and previously unknown attribute combinations
2) Use the program to also rapidly do mass testing of individual players for the real player ranking list
3) Use what turns up out of it all to make better informed guesses for the selection of weights to try and match real player rankings
Attribute testing isn't the main benefit of LightingFlik's tool IMO, it's stuff like what's he posted finding about how the game works, and the ability to do mass tests of player swaps with ease, and perhaps a lot more.
The ability to do attribute testing 20+ times faster is nice, but not exactly a game changer. You could already do it before, and it was done - by FM Arena, Harvest Green, myself, etc. - it's just about how much time are you willing to sink into it.
That said, the ability to do 1000 samples in 3 minutes or so does enable some previously unrealistic things, like getting accurate read on the GK attributes. Even my own manual findings on GK have not been that satisfactory.
I did do agility, I forgot to add it in. I'll add it in now.
Edit: Actually I see why I left agility out. I've got mixed results on it. What we can glean from that at least is that agility probably doesn't matter that much, but I want to do some more testing on it to provide further clarification.
Here's what I have so far:
Outfield:
Pace = 100.00
Acceleration = 93.84
Dribbling = 25.28
Vision + Passing + Decisions combo = 23.58
Strength = 15.72
Marking + Tackling + Positioning combo = 15.23
Crossing + Technique + Longshots combo = 11.23
Leadership = 4.67
Aggression = 1.62
GK:
Reflexes = 19.88
Acceleration = 12.10
Pace = 12.10
Command of Area = 8.53
Determination = 3.97
One On Ones = 2.31
Aerial Reach = 2.19
First Touch = 2.19
Rushing Out (Tendency) = 1.74
Punching (Tendency) = 1.70
Concentration = 1.46
Technique = 0.41
Handling = 0.20
Throwing = -0.24
Communication = -0.45
Eccentricity = -2.55
Jumping Reach = -3.63
Strength = -4.11
Kicking = -4.41
Stamina = -6.70
Notes:
The attributes are represented as relative to each other. +1 reflexes on the GK is equivalent to +0.2 pace on ALL outfield players.
I count +1 for win and +0.333 for draw (before conversion), that's all I'm judging by.
The negative GK attributes are not mistakes from one-off outliers. For instance for Strength here's the tests and trend:
Strength 1 (1000 samples) = W 262 (26.2%), D 209 (20.9%), L 529 (52.9%), 1.29-1.85
Strength 10 (1000 samples) = W 257 (25.7%), D 203 (20.3%), L 540 (54.0%), 1.27-1.84
Strength 12 (1000 samples) = W 252 (25.2%), D 212 (21.2%), L 536 (53.6%), 1.27-1.86
Strength 16 (1000 samples) = W 244 (24.4%), D 204 (20.4%), L 552 (55.2%), 1.23-1.83
I'm being selective about what I test. Some I just do '1' and '18'. If the result is unusual, like Strength, or crucial, like Reflexes, I do more intervals. I haven't done this for 1 or 2 yet though such as GK stamina.
Because attributes are non-linear, I've decided to represent the results here as more typical/realistic values rather than simply '1' vs. '20'. A few examples to illustrate what some of the above figures are derived from:
Pace 15 > 16 or 14 = 51.4165
Acceleration 15 > 16 or 14 = 48.25
Leadership 8 > 18 = 24
Strength 6 > 10 = 32.334
Aggression 6 > 12 = 5
Dribbling 6 > 10 = 52
I've found a few (relatively minor but notable) differences to blackbird's results so far. For instance aggression does matter, but mostly only going from 1 > 6 (blackbird only did 4 > 8, and higher). I did not see a significant difference for Strength going from 10 > 20, but 6 > 10 is significant. I'll be posting the extended results later that bears this out.
It could partly be that blackbird used a tactic that affects the outcomes. I've used a variant of the Mars Attacks V7 tactic (433 formation) for my testing. I've previously done all my testing with the Knap tactic, so when I realized I was using the V7 tactic, I figured I'll just continue with it anyway since that will shed light at the same time on whether tactics/formations affect attribute outcomes or not. And it doesn't seem that they do.. so long as they are an intense gegenpress style tactic at least I guess.
Later on my results should be able to mostly complement blackbird's results to give a more complete picture (i.e. values for 1, 4, 6, 8, 10, etc.)
It's ok, I think mainly I just need to take a break and then understand properly how to use it. I had that crash issue, but I can mostly avoid this so it's not really that.
If no one else was doing the tests, I'd probably push myself to do it, but since you & others are getting results (and more sophisticated than what I can currently do), no point putting myself through tedium. The benefits of this are so good that I will get back into it after a brief break though.
What made me call it a day is I edited the attributes in the JSON as you suggested, creating a changed JSON and a control JSON which did get different results, but then when I edited the control JSON it kept producing identical results when it shouldn't. And the problem is, I don't know if this is you're using the wrong methodology (i.e. maybe you don't realize its already locked in certain attributes at this point), or if I'm doing something wrong. I suspect it's the latter, but I can't tell. So I really need to develop my understanding first.
It did occur to me writing my last post that the program can also be applied to the real player alignment process too, which will speed it up in the same way it sped up attribute testing.
I've done an outfield squad of Haalands btw, with manual testing. Scores around +150 g/d in the Premier League with Knap tactic. For reference, the max possible is somewhere around +450 g/d.
LightningFlik's program is giving us more precise data to use, but so far it doesn't look like much needs to change.
Any new changes probably still need to go through the same process of lining up performances of real players correctly and deriving weights from that. Having more precise attribute data is going to improve this, because it will narrow down the possible combinations that explain things, but this manual adjustment to line up real players still has to be done using the data.
A change to the GK weights might happen quicker though, as I think it's likely to have a few errors that will be uncovered through this new attribute testing method.
---
Is anyone doing the GK attributes?
I'm still having some troubles with getting the testing working properly or easily, and I'm quite burnt out by fiddling around with it. So I probably won't be contributing much immediately, especially with others having a better go of it than me.
Looking forward to more insights from others
Pace: -2.984 / -1.608 / -0.736 / -0.074 / +0.614 / Best: 20 / Range: 3.598
Acceleration: -2.496 / -1.430 / -0.522 / +0.214 / +0.726 / Best: 20 / Range: 3.222
Work Rate: -0.874 / -0.090 / +0.182 / +0.254 / +0.302 / Best: 20 / Range: 1.176
Agility: -0.538 / -0.246 / +0.054 / +0.164 / +0.388 / Best: 20 / Range: 0.926
Dribbling: -0.126 / +0.006 / +0.266 / +0.498 / +0.762 / Best: 20 / Range: 0.888
Jumping Reach: -0.292 / -0.256 / -0.028 / +0.250 / +0.596 / Best: 20 / Range: 0.888
Concentration: -0.304 / +0.008 / +0.190 / +0.282 / +0.476 / Best: 20 / Range: 0.780
Balance: -0.328 / -0.108 / +0.060 / -0.002 / +0.360 / Best: 20 / Range: 0.688
Stamina: -0.314 / -0.102 / -0.048 / +0.030 / +0.260 / Best: 20 / Range: 0.574
Anticipation: -0.278 / -0.074 / +0.068 / +0.216 / +0.268 / Best: 20 / Range: 0.546
Finishing: +0.104 / -0.142 / +0.036 / +0.360 / +0.392 / Best: 20 / Range: 0.534
Strength: -0.102 / +0.048 / +0.138 / +0.166 / +0.370 / Best: 20 / Range: 0.472
Important Matches: -0.054 / +0.034 / -0.028 / +0.294 / +0.396 / Best: 20 / Range: 0.450
Dirtiness: -0.018 / +0.114 / -0.092 / -0.096 / -0.294 / Best: 8 / Range: 0.408
Composure: -0.096 / +0.036 / +0.218 / +0.216 / +0.306 / Best: 20 / Range: 0.402
Determination: -0.182 / +0.034 / -0.060 / +0.198 / +0.104 / Best: 16 / Range: 0.380
Crossing: -0.080 / +0.076 / +0.160 / +0.288 / +0.244 / Best: 16 / Range: 0.368
Decisions: +0.190 / -0.096 / +0.054 / +0.254 / +0.178 / Best: 16 / Range: 0.350
Long Shots: -0.084 / +0.086 / +0.036 / +0.056 / +0.250 / Best: 20 / Range: 0.334
Consistency: +0.000 / +0.024 / +0.080 / +0.224 / +0.276 / Best: 20 / Range: 0.276
Heading: -0.014 / +0.128 / +0.034 / +0.144 / +0.238 / Best: 20 / Range: 0.252
Vision: -0.014 / -0.086 / +0.104 / +0.054 / +0.162 / Best: 20 / Range: 0.248
Teamwork: -0.062 / +0.042 / +0.060 / -0.026 / +0.150 / Best: 20 / Range: 0.212
Aggression: +0.094 / +0.056 / -0.044 / +0.042 / +0.148 / Best: 20 / Range: 0.192
Flair: +0.106 / +0.120 / +0.024 / -0.068 / +0.118 / Best: 8 / Range: 0.188
Marking: -0.008 / +0.088 / +0.118 / +0.132 / +0.178 / Best: 20 / Range: 0.186
Positioning: -0.014 / +0.032 / +0.024 / +0.058 / +0.156 / Best: 20 / Range: 0.170
Technique: +0.130 / +0.018 / +0.064 / +0.112 / +0.188 / Best: 20 / Range: 0.170
Leadership: -0.020 / +0.004 / +0.058 / -0.042 / +0.128 / Best: 20 / Range: 0.170
First Touch: -0.038 / +0.072 / +0.112 / +0.090 / -0.002 / Best: 12 / Range: 0.150
Injury Proneness: -0.010 / -0.006 / +0.088 / +0.114 / +0.010 / Best: 16 / Range: 0.124
Passing: +0.050 / +0.096 / +0.136 / +0.148 / +0.154 / Best: 20 / Range: 0.104
Bravery: +0.022 / +0.056 / +0.084 / +0.114 / +0.048 / Best: 16 / Range: 0.092
Off the Ball: +0.164 / +0.094 / +0.184 / +0.100 / +0.176 / Best: 12 / Range: 0.090
Tackling: +0.062 / +0.110 / +0.020 / +0.038 / +0.102 / Best: 8 / Range: 0.090
Versatility: -0.044 / +0.018 / -0.016 / +0.002 / +0.000 / Best: 8 / Range: 0.062
Natural Fitness: +0.000 / +0.000 / +0.000 / +0.000 / +0.000 / Best: 4 / Range: 0.000
This test shows that some attributes are indeed non-linear
This all just about lines up with what I found through manual testing, so there's not much to change in my ideal template I think. I didn't have this level of precision though of course.
There's a few small surprises from the new data so far:
Aggression (apparently useless)
Strength (at 20)
Finishing (at 16-20)
Dribbling (at 4-8)
Dirtiness (seems to have an unusual mechanic, I wouldn't go by the results here for this one)
Vision (at 4)
Determination (at 8)
I kind of already knew that strength and finishing had benefits at higher levels, I surmised that it's not worth the extra CA, but I thought it was more linear than what the results show here. I still don't think they're worth getting though. Perhaps players who happen to have very high finishing are more worthwhile though.
Determination seems to work not too bad at lower levels, but there's also the effect on training to consider, so I think higher is still better.
If we surmise pace as the difference between 12 and 20 pace, it's 1.35/8 = 0.16875 per 1 pace
So
20 strength = 2.19 pace
20 finishing = 2.32 pace
20 longshots = 1.48 pace
20 agility = 2.30 pace
20 balance = 2.13 pace
Strength, agility and balance are very similar, but agility has a much more severe impact when 8 or lower. Balance is inbetween, which is a little surprising, because my testing found balance to be a bit more important than agility.
Crossing seems to carry marginal improvements up to ~16. Crossing 8 > 16 is equivalent to +1.70 pace. For comparison, Strength 12 > 20 is +1.37 pace, so that's a considerable benefit. Other attributes are more important of course. Consider 8 > 12 composure for instance, which is +1.29 pace for half the attribute points (so +2.58 pace for composure vs. +1.70 pace for crossing).
I did a few more tests yesterday, which I'll post since they include temperament and professionalism which blackbird hasn't included in his test:
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Teamwork 14 Meta vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 52 (17.33%), Loss 31 (10.33%), avg. 2.47-0.88
Temperament 20 Meta vs Man Utd (Home) Fully healed 300 sample = Win 222 (74.00%, Draw 43 (14.33%), Loss 35 (11.67%), avg. 2.44-0.82
Professionalism 8 Meta vs Man Utd (Home) Fully healed 300 sample = Win 220 (73.33%), Draw 51 (17.00%), Loss 29 (9.67%), avg. 2.50-0.89
Natural Fitness 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 213 (71.00%), Draw 56 (18.67%), Loss 31 (10.33%), avg. 2.36-0.82
Determination 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 192 (64.00%), Draw 60 (20.00%), Loss 48 (16.00%), avg. 2.39-1.00
98con/97fit/75morale Meta vs Man Utd (Home) Fully healed 300 sample = Win 186 (62.00%), Draw 68 (22.67%), Loss 46 (15.33%), avg. 2.16-0.92
So it seems that temperament 20 might not be worse than temperament 10 as Orion found. It doesn't seem to have any benefit over 10 either though.
Professionalism surprisingly seems to have little downside going from 13 > 8. Of course it matters for training and other things.
The natural fitness result seems to align with blackbird's result that it makes no difference. But it is nonetheless very important because it affects match sharpness loss rate outside of the matches, which is very impactful on performance.
Even a slight change in fitness and/or morale has a significant impact on results. So trying to get that extra ~2% condition or little boost to morale prior to a match really does matter. We see here -2% condition & -25% morale is worse than 14 > 1 Determination for the entire outfield.
LightningFlik said: 100% home fans: home 588 / draw 181 / away 231, avg 3.38-2.36
100% away fans: home 234 / draw 173 / away 593, avg 2.18-3.19
50%/50% fans: home 399 / draw 191 / away 410, avg 2.72-2.73
No fans: home 407 / draw 199 / away 394, avg 2.70-2.74
75/25 fans: home 389 / draw 195 / away 416, avg 2.72-2.79
75,000/25,000 fans: home 512 / draw 202 / away 286, avg 3.08-2.46
(These last two tests were to see if the absolute value of fans made the positive impact or the relative strength. Given the absolute number of fans makes a difference then there are performance gains in upgrading your stadium as much as possible, since every fan counts.)
Bruh what
Fan count affecting results was not something I was expecting
I guess this has implications for team selection when playing
(And, if I am, I can bake in the recalculation.)
I'm using a realtime editor to change the attributes in the game before exporting to JSON
I haven't taken the time yet to examine how your program works. So I don't know what you mean by 'fixture object'. Are you saying we can just edit the attribute data in the JSON file after export?
Let me know the process, because having to change the CA for each player is a real pain
If I change an attribute that has CA weight, I have to change the player's CA to the new Recommended CA, otherwise it messes with the results.
I had my fingers crossed initially that the game would automatically just recalculate the CA for the player, but its not the case. It reduces every other attribute a little to make up for it.
When I added the correct CA, I got results that lined up with expectation (i.e. teamwork being marginally beneficial, rather than 8 > 14 teamwork producing significantly worse results)
Dominating means winning almost all games with ease, which we measure with the points tally.
There is a high correlation with fitness and physical attributes. We don't know what it's specifically composed of, yes, but I wouldn't go so far as to say it's not measurable.
Jolt said: What @LightningFlik mentioned was that certain groups of attributes may be made more effective by an external attribute threshold (Concentration >11).
I think you've made an error of interpretation here. I think LightningFlik is saying that concentration >11 was necessary for certain miniscule match focus boosts, not for groups of attributes to kick in or be more effective.
It's possible similar things are going on with attributes, we don't have anything to suggest that yet.
One thing that suggests its not going to be like that is that I scaled down my template for lower leagues, and it seemed to work just as good. So I think it's about attributes relative to the opposition, rather than set thresholds for attributes to take effect.
Jolt said: Of course, this is an utterly gargantuan amount of work. All visible attributes, plus the four hidden ones we know have an effect on performance (Consistency, Important Matches [this one would require editing match reputation or editing rivalry between the teams], Pressure, Dirtiness), and we're talking about close to 10,000 3-attribute combinations.
Even testing just 3 variations per attribute (7, 12, 17), and running 10 games per variation to get a sample size, and we're talking about hundreds of thousands of games.
So it is understandable that noone followed up on it.
I think it's viable to do in a certain way
We can get rid of some attributes such as long throws, penalty taking, etc.
Then we can also discard some combinations such as vision + natural fitness + technique. And order of the combination doesn't matter either.
I did about 11,000 samples last night. Not sure on the exact numbers, but this means our processing capacity just increased at least 20-fold. Maybe applying AI to it can strip away the remaining problem of tedium.
I can't remember what I got this from, but I think combos can supposedly be up to 5 attributes.
What's holding me back right now is that it's super tedious to adjust CA in line with attribute changes in the realtime editor. Maybe the official realtime editor automatically does this, might see if I can use it later.
- In possession: Off the Ball is rolled twice every 15 seconds (instead of just once) when your team has the ball and the higher of the two values is used.
- In Transition: Teamwork is re-rolled after possession change and the higher of the two values is kept.
- Out of Possession: Marking is rolled twice while the opponent has the ball and the higher of the two values is used.
- Attacking corners: Anticipation is rolled twice when you have a corner and the higher result is used
- Defending corners: same as above except when the opponent gets a corner
- Attacking free kicks: same as above
- Defending free kicks: same as above
- Throw-ins: Re-rolls anticipation or gives a +1 bonus to the original value if it was already better
- Penalties: a flat +1 bonus when taking a penalty
And by the way, I think that to get these bonuses, a player's current Concentration roll has to be >11, which means players with a low concentration attribute never see these buffs.
So it sounds like match focus is kind of like a consistency modifier for those particular attributes you mention. Sounds like 'transition phrase' would be the one to pick, since teamwork actually does a little something.
Jolt said: That leaves me wondering why didn't @Zippo find any correlation between them: https://fm-arena.com/find-comment/40372/ in their testing. As I mentioned back then, those findings were extremely surprising.
I think I can shed some possible light on this. My own testing has shown that there are certain combinations of technical attributes (i.e. crossing + long shots + technique) that produce a significant difference, but this only came through when the team was dominating.
I think what this means is that when Man City play Barrow in the FA Cup, they're going to go above and beyond what Burton would do to Barrow, but only so long as their physicals mean they can keep it up - and in this case, they should have enough physical superiority to keep it up all game, whereas Burton might have to call it a day at 5-0 60 minutes in.
But if its say Man City vs. Luton, Man City should start with enough physical advantage to execute all that technical finesse for a while, but then I think the idea is that it ends up becoming a slogfest for much of the match where whoever makes the least serious mistakes (i.e. low dirtiness, high concentration) wins the battle.
IMO I think it's deliberate but probably unintentionally unbalanced, because it's common sense that if a player is worn out, he's not going to be pulling off clever technical skills.
With these combinations not showing up in tests, it would likely be because these tests are controlled where all players about equal in fitness all through the match due to near identical attributes, healing, etc.
I cumulatively improved it from W 132, D 36, L 32, avg. 2.02-0.94 to W 142, D 39, L 19, avg. 2.21-0.79.
For comparison, the Knap 424 scored W 135.2, D 39.2, L 25.6, avg. 2.36-0.86.
I then tested it with Rushall, finished 3rd. I thought a bit disappointing, but not too bad.
I tried Luton next. I got 1st with an unselected team straight off the bat. I thought bloody hell, this is it. To get 1st with Luton with a selected team is very rare. To get it with an unselected team is unprecedented. I think I was even a few points clear of 2nd.
But then I got 6th and 10th. And with Man City, the results were 1st but only 93pts and 100pts.
But today I see that I accidentally used what scored W 135, D 39, L 26, avg. 2.35-0.89 instead.
After fixing it up, I've still only got 7th as unselected Luton and 4th as selected Luton as the two initial samples.
So perhaps tactics are going to be somewhat safe from this brute-forcing method. This and the dirtiness results suggests to me that the non-linear nature of attributes and their interaction with external factors such % condition, means that we can't simply universalize a tactic that works perfectly for one team. Each team would have a slightly different optimum, and the method seemingly can't really tell you what tactic works best with most teams overall.
But I guess what I can try next is to brute force Luton rather than the meta team, and see what that produces.
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Dirtiness 12 Meta vs Man Utd (Home) Fully healed 300 sample = Win 209 (69.67%), Draw 54 (18.00%), Loss 37 (12.33%), avg. 2.39-0.94
Dirtiness 12 Meta (1 day before) vs Man Utd (Home) Fully healed 300 sample = Win 197, Draw 67, Loss 36, avg. 2.33-0.91
We see above that Dirtiness 8 > 12 seems to make no difference when the squad is set to 100% fitness.
Longshots Technique Passing Crossing 16 vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 98, Loss 110, avg. 1.35-1.49
Dirtiness 12 Longshots Technique Passing Crossing 16 vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 79, Loss 129, avg. 1.35-1.62
By contrast, when fitness is left to standard levels - around ~90% condition ~100% match fitness on average in this case - there does seem to be a significant difference.
This will require further testing to clarify. Other attributes aren't working this way, and are producing results in line with expectations.
I believe that the new method I'm using has fixed up my previous teamwork & technicals results, which should be disregarded as erroneous. I think it's mainly that I didn't adjust CA upwards in the realtime editor to match the increases, which I've done accurately now. I'll probably edit the previous post later to remove the old results to avoid confusion.
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
Control Meta (Knap 424 tactic) vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 44 (14.67%), Loss 39 (13.00%), avg. 2.35-0.83
Teamwork 14 Meta vs Man Utd (Home) Fully healed 300 sample = Win 217 (72.33%), Draw 52 (17.33%), Loss 31 (10.33%), avg. 2.47-0.88
So 8 > 14 Teamwork is now seemingly marginally beneficial. This is what I would have expected.
Pressure 1 Meta vs Man Utd (Home) Fully healed 300 sample = Win 146 (48.67%), Draw 73 (24.33%), Loss 81 (27.00%), avg. 1.60-1.00
Thought I'd try a very obvious one just to make sure things are on the right track and provide a yardstick to compare these lesser attributes to. Big drop from Pressure 13 > 1 as expected.
Longshots Technique Passing Crossing 16 Meta vs Man Utd (Home) Fully healed 300 sample = Win 236 (78.67%), Draw 48 (16.00%), Loss 16 (5.33%), avg. 2.88-0.84
Control Meta (Knap 424 tactic) vs Wolves (Away) Fully healed 300 sample = Win 256 (85.33%, Draw 31 (10.33%), Loss 13 (4.33%), avg. 2.75-0.83
Longshots Technique Passing Crossing 16 Meta vs Wolves (Away) Fully healed 300 sample = Win 274 (91.33%), Draw 21 (7.00%), Loss 5 (1.67%), avg. 3.09-0.82
Longshots Technique Passing Crossing 16 Meta vs West Ham (Away) ~90% condition 300 sample = Win 92, Draw 98, Loss 110, avg. 1.35-1.49
These attributes performed a quite a bit better than I expected. I did find in my testing that this particular combination yields significantly higher goal difference at very high performance levels, but this is showing significant benefits even at a lower performance level - albeit at a cost of +32 total attribute points and a good deal of CA to match.
Control Meta (Knap EF 424 IF HP V2 P101 AC) vs Man Utd (Home) Fully healed 1000 sample = Win 676 (67.6%), Draw 196 (19.6%), Loss 128 (12.8%), avg. 2.36-0.86
V7 Variant B vs Man Utd (Home) Fully healed 1000 sample = Win 656 (65.6%), Draw 221 (22.1%), Loss 123 (12.3%), avg. 2.14-0.86
424 Mountain King vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 208 (69.33%), Draw 63 (21.00%), Loss 29 (9.67%), avg. 2.36-0.78
424 TON 424 HP HUB MC P105 EFC vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 221 (73.67%), Draw 50 (16.67%), Loss 29 (9.67%), avg. 2.15-0.75
424 'Park the Bus' (default tactic) vs Man Utd (Home) Fully healed 300 sample (not full familiarity) = Win 50 (16.67%), Draw 78 (26.00%), Loss 172 (57.33%), avg. 1.37-0.66
I think an issue with comparing tactics is that the players you are fielding matter, and in real play you're more likely to get good players in certain positions rather than others.
So we can say, this tactic does best in a controlled test, and that tactic does best in a default plug-and-play with Luton, but it would be difficult to find conclusively that X tactic is 2% better than Y tactic. But you could demonstrate it's 2% better in 5 or 6 scenarios say.
Meta team Teamwork 14 (from 8) Fully healed 200 sample = Win 107 (53.23%), Draw 50 (24.88%), Loss 44 (21.89%), 1.93-1.00
Meta team Determination 8 (from 14) Fully healed 200 sample = Win 112 (55.72%), Draw 44 (21.89%), Loss 45 (22.39%), 1.93-0.96
Control reference: 63.0% Win, 24.2% Draw, 13.0% Loss
The Teamwork result could be wrong, as with other CA-changing attributes I've done. I will redo all these with a better method and proper CA edits. But from this result, it suggests Teamwork is a negative attribute.
The non-CA attribute results should all be about right though, which includes Determination.
A little surprisingly lower determination isn't too bad, at least in this range. 14 > 8 Determination is equivalent to 8 > 12 dirtiness.
Since no one has answered you, I'll say no way that I know of, but perhaps there's a way of editing the skin to do it.
Kma said: Helllo please @GeorgeFloydOverdosed is there some sort of hidden attribute or stat that will determine players to become staff or not or it's totaly random? have you test that?
and what that numbers meaning in the pre game editor? 95% of players are like that?
I haven't tested this. There's probably a few factors involved. That section in the editor is I think just how willing they'd be to accept the particular role if you make them an offer, but there would have to be other stuff influencing them to become staff member instead of retire from game entirely, and so forth.
LightningFlik said: Pretty sure these relationships have no effect in matches by the way.
I've found the data that represents them in memory: it tracks games and minutes played together (plus two other values I've not deciphered yet). But I can't see these in the fixture data, meaning they don't factor in to on-field performance.
Yeah I mentioned this in my tactic test findings, so I can confirm this would be true
LightningFlik said: Good question. I ran my experiment using on tiniest database, EPL only, managing default Aston Villa, using Knap's 424 tactic, playing a friendly at the earliest opportunity against Bristol City. (The exact lineup I used probably doesn't matter, but suffice to say I let the assistant pick everything.)
You could recreate those conditions if you wanted to double check my result.
If you match my result then the discrepancy might come from a quirk of your meta-players' attributes. If you don't match my result then I can re-run my experiment later. Maybe I had incorrectly mapped the anticipation attribute in my test or maybe I was still simulating games incorrectly (for a time the player's manager would make no subs).
Also you could pastebin me your fixture JSON and I could re-run my experiment on that and see if I get the same result as you.
I thought the tactic could be part of it, but apparently not, as we've used the same tactic.
If it was pace/acc, then it being a quirk of the meta player's attributes would make sense - as pace/acc is on a cliff edge for the meta team, but anticipation shouldn't be like that as it's fairly linear I think. But I did check Aston Villa's players because maybe they were all going from '16' > '11', where '11' wouldn't be much worse probably. But I checked a few and they were mostly ~14, so it would be '14' > '9', which is similar to what I did.
So I reckon it's some methodology issue. Main thing is it doesn't seem to be an issue with the program now.
Take that one with a grain of salt right now. It could be true, or it could just be my realtime editor stuffing things up.
Meta team Pace/Acceleration 13 (from 15) Fully healed 200 sample = Win 60 (29.85%), Draw 43 (21.39%), Loss 98 (48.76%), 1.23-1.62
So:
-10 Aggression = No effect
-8 Leadership = -2% win%
+4 dirtiness = -8% win%
-4 pace/acc = -33% win%
So if pace is 100 weight, then dirtiness 24 and leadership 3.
These aren't conclusive at all, just interesting to see what the rough picture is first from this new method.
Meta team Aggression 1 (from 11)(Leadership 8) Fully healed 500 sample = Win 308 (61.48%), Draw 108 (21.56%), Loss 85 (16.97%), avg. 2.09-1.05
Meta team Anticipation 8 (from 13)(Leadership 8) Fully healed 200 sample = Win 84 (41.79%), Draw 52 (25.87%), Lose 65 (32.34%), avg. 1.51-1.25
So we see here that Aggression doesn't seem to do anything. If anything, a very slight gain from lower aggression.
The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier.
It's also surprising to me how strong the RNG is. Even comparing 100 match samples to 500, there's a 10% difference. Though perhaps I need to adjust the way I'm testing for a more controlled result.
LightningFlik said: It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game?
I was going to say maybe Leadership is keeping morale of other lesser players high, but then I realized that that's one thing that can probably be ruled out, as I'm giving them all identical attributes and leadership.
And the same could be said for the 'leader' boosting other's determination, etc. to his level.
Vs. Nottm Forest (Man City away) (meta team, but with 16 leadership - forgot to change that):
Meta team Dirtiness 12 = Win 45, Draw 26, Loss 30, avg. 1.55-1.18
Meta team Dirtiness 10 = Win 38, Draw 28, Loss 35, avg. 1.45-1.27
Meta team Dirtiness 9 = Win 43, Draw 33, Loss 25, avg. 1.51-1.02
Meta team Dirtiness 8 = Win 50, Draw 34, Loss 17, avg. 1.82-1.13
Meta team Dirtiness 7 = Win 45, Draw 27, Loss 29, avg. 1.64-1.27
Meta team Dirtiness 5 = Win 41, Draw 28, Loss 32, avg. 1.48-1.17
Meta team Dirtiness 1 = Win 44, Draw 24, Loss 33, avg. 1.44-1.07
All of those are 100 samples each and use standard fitness, morale, etc. values.
Meta team Dirtiness 8 Fully healed 100 sample = Win 69, Draw 15, Loss 17, avg. 2.20-0.76
Meta team Dirtiness 1 Fully healed 500 sample = Win 321, Draw 107, Loss 73, avg. 2.06-0.89
Meta team Dirtiness 8 Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
Meta team Dirtiness 12 Fully healed 500 sample = Win 276, Draw 119, Loss 106, avg. 1.97-1.06
As percentages:
Dirtiness 1 = 64.2% Win, 21.4% Draw, 14.6% Loss
Dirtiness 8 = 63.0% Win, 24.2% Draw, 13.0% Loss
Dirtiness 12 = 55.2% win, 23.8% Draw, 21.2% Loss
In the 100 sample tests, 8 dirtiness performs best. In the 500 sample one, 1 dirtiness just about equals 8 dirtiness. 12 dirtiness is clearly worse than 8 dirtiness in both.
I can't recall what testing I did exactly which led me to believe this, but I think I have found at some point that 1 dirtiness did worse than 8 dirtiness, so I did figure we have a kind of bathtub curve going on with dirtiness. But on the other hand, this leaves players like Maeda inexplicable. So I ended up just setting it aside but having it in the back of my mind, and it's one reason I wanted to test dirtiness first.
And so I'm not surprised by these results showing 8 is optimal, though I would have expected it to be 6 rather than 8.
What's confusing though is that in the larger sample with perfect fitness/morale/etc 8 does the same as 1. Maybe the effect of dirtiness changes according to fitness condition of players involved say. Or maybe 8 really is the same as 1 and there was too much randomness in the fewer samples.
Thinking about it.. Maeda has really high pace/acc, so maybe he benefits from avoiding doing dirty things, because then he takes advantage of his high pace/acc instead. But if a player has low pace/acc, or is tired, he's better off doing something dirty to get something out of the game. Maybe that's what going on.
In any case, the data does seem to support the fact at least that lower dirtiness = generally better results, with distinguishable significance. We'll see how it compares to another attribute. I think I'll test leadership next.
Edit:
Meta team Leadership 8 Fully healed 500 sample = Win 307, 105, Loss 89, avg. 2.01-1.00
Compare to:
Meta team Dirtiness 8 (Leadership 16) Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
So +8 leadership seems to have slightly better results. About half the difference of dirtiness 8 > 12. So 2 dirtiness = 8 leadership? Will test aggression next to compare.
Before this I did test Leadership again and got results of 3rd, 4th, 5th with Leadership @ 16 instead of 8. So it seems like leadership does have some effect. Not sure how I missed it, I probably only tested it once, got a 5th (which means no position change) and moved on. My guess now is it would have a weight of ~10 maybe.
Needed to use ././ to get it working
But then also:
# ././fm24ctl get-next-fixture next-fixture.bin
fm24_harness_p10.dll not found; injecting fm24_harness.dll (it must be a protocol 10 build)
injected into fm.exe (pid 22960); see harness.log for the build check
the harness did not start in fm.exe (wrong game build?); see harness.log
Log:
10:55:26.461 loaded into fm.exe at 0000000140000000
10:55:26.461 refusing: fm.exe timestamp 1744804949, expected 1744804480 (different game build)
I'm using the Epic Games version, is this why?