hziz1 said: TOP 3 attributes is PACE, ACCELERATION,JUMPING REACH Expand
Don't forget Dribbling. Jumping Reach (71.6 points) and Dribbling (69.8 points) are pretty close together. And then comes Balance (65.3) which isn't negligible. And then comes Concentration (64.5 points) and then Anticipation (64.3 points) and so on. https://fm-arena.com/table/26-player-attributes-testing/
Do you see the point? I can create a player which has e. g. lower Jumping Reach but higher Dribbling, Balance and Concentration and is therefore a better player. It is good that we have found Meta-Attributes but we should see these attributes in the right context and not in a "I don't care all other attributes" attitude because this is not adequate. Winning a league with only a few player attributes is spectacular, but it is not the optimal. Winning a league against AI managers isn't difficult. I can do this with any average team when I use the right tactic and improve team cohesion. That can't be the benchmark.
Cherknam said: I'm confused. How can the minimum RNG score for the 3.0 test be higher than the maximum RNG score for the 2.0 test, if the match engine and the tactic are the same? Surely at least one of these RNG calculations is incorrect. Expand
There are a) changes and b) never forget: There isn't a minimum RNG. It is always a confidence interval like "Minimum score with 90% confidence or 95% confidence etc." so you can get a higher or lower score outside the confidence interval.
ZaZ said: The best moment to use substitutions is at half time. Expand
Then you forego the increase in stamina at half time. If you assume that the starting eleven are better than the bench players, then you should change them after half-time.
It's really strange when you say "Only 9 attributes matter but I set other attributes to 13 because these attributes matter too".
I've been testing a lot lately and have regained the joy of FM (I haven't bought an FM every year in the last few years because it's become too easy for me). But the results of tests should always be presented objectively and the Reddit user didn't do that. All attributes matter and it is interesting to find out how important each attribute is for each position. I would beat e.g. this reddit users team with my 19 acceleration, 19 pace, 19 anticipation, 19 jumping reach, 20 finishing, 20 technique, 20 first touch, 20 heading striker. This problem is more complex than reddit users posts. Yes, we found meta-attributes but this isn't the end of the research.
MeanOnSunday said: There are simplistic problems where it’s possible to isolate a single variable, but FM is really not such a case. Expand
We can isolate single variables in FM. Any attribute is a variable. Luckily, FM is just a game and there is no reason to assume that a higher attribute value leads to worse results (regarding your glue example). When I read hints of the developers I always read things like this: "A goalkeeper's ability to save penalties is initially governed by his Anticipation, Reflexes, and Concentration in reacting to the moment the ball is struck. His Acceleration will help his immediate chances of reaching the ball successfully, whilst his Agility, Reflexes and Handling will ultimately determine whether or not he pulls off the save." So a higher value is always better and maybe all these attributes are additively linked like e.g. "0.4*anticipation+0.8*reflexes+0.3*concentration+0,2*acceleration+0.4*agility+0,3*handling" when there are penalties. We shouldn't forget that this is a game designed for entertainment. But these are all theories and assumptions. Back to practice:
I can beat machine learning with our single variable approach and I can prove it:
Spoiler
I will make a team which is worse than the opponent team according ykykyk rating. ykykyk is the result of machine learning and ykykyk is based on 40 million matches. I will even use the tactic which the machine used for the 40 million matches (ZaZ - Blue 3.0). My team has only one information source: A few thousand matches on fm-arena tables which lead to the conclusion that pace and acceleration are the most important variables. So it is me against the machine:
Both teams have this starting players, bench players and tactic:
My Team is "Arsenal 23/24" and my opponent is "Arsenal 23/24 (2)". ykykyk ratings are higher for all "Arsenal 23/24 (2)" counterparts on the field and on the bench (except goalkeepers):
So "Arsenal 23/24 (2)" should beat my team according machine learning. These are the results after 40 matches:
"Arsenal 23/24" : "Arsenal 23/24 (2)"
This is a clear result beyond RNG. It isn't surprising when you consider how strong pace and acceleration are on the fm-tables. 40 million matches aren't enough for a machine to see this. This is the big advantage when you just change one variable to see what are the really important variables. Machine learning is a great tool when you are Google and have the ressources to simulate billion of billions of matches. We don't have these ressources so our approach is better than machine learning with 40 million matches.
MeanOnSunday said: When you test attributes one by one and for all positions it can only tell you if an attribute is important for the majority of positions. Expand
I agree with you.
MeanOnSunday said: The only way to know at a single position level is vary multiple attributes in the same experiment and vary by position as well (this is in effect what the machine learning data is reporting). Expand
I don't agree. You have to vary one attribute for one position to know its value for this position (in a particular tactic). If you vary multiple attributes than you don't know which attribute is important for the performance increase.
Spoiler
Why do scientists change only one variable in a controlled experiment?
In a controlled experiment, scientists change only one variable at a time in order to accurately determine the effect of that specific variable on the outcome. By isolating and manipulating one variable while keeping all others constant, scientists can establish a cause-and-effect relationship between the variable they are testing and the results they observe. This helps in drawing accurate conclusions and making reliable predictions based on the experiment's findings. If multiple variables were changed at the same time, it would be difficult to attribute any observed effects to a specific variable, leading to unreliable or inconclusive results.
ZaZ said: Just as a reminder, the fact that one attribute did not get impressive results in the table of FM-Arena does not mean that attribute is useless. For example, stamina might be impactful only for two or three players, like DM and WBs, which represents less than 30% of the team. In that case, even if Stamina is as impactful as Pace and Acceleration for those positions, it will still give worse results than attributes that have impact across all roles. Expand
Decision is for every position more expensive than Anticipation (except SC) but the results on the tables are worse in comparision to Anticipation (FM22/FM23/FM24). Maybe there is one position (or even two positions) where Decision has a better cost-benefit ratio than Anticipation. But that doesn't matter. Why? Because you can't train Decision just for one position. You always must train Decision for all positions:
So a high value for Decision in a training session is always a bad thing.
Yarema said: Do you use individual training? Because that's where I think match practice shines, focusing on those key attributes and possibly new position training. Expand
Individual training has no effect on attributes within a position (no difference between e.g. no-nonsense centre-back and ball playing defender). It has an effect when you train a new position e.g. right defender learns wing-back position. But this is valid for all training sessions.
My test: Spoiler
DL is trained as DL (Wing-back attack) DR is trained as WB (Wing-back attack) Only match practice training Starting point of all attributes = 50 (ingame = 10) Players are 20 years old 4 runs
It seems that DR doesn't train natural fitness and strength.
roman said: This is probably the fastest way to grow CA overall, but I think you will get higher quality players if you can focus the attribute growth on highly weighted attributes. Expand
There is a big problem. One example: Anticipation is good but Decision is bad in context of CA efficiency according to the test results on fm-arena. But there is no training session which trains more Anticipation but less Decision.
Excerpt from MaxEBFM results:
Another example is Acceleration/Pace (both are good) and Agility (bad). I tried to beat the system by using a good and a bad coach but that doesn't work either. Anticipation/Decision is trained by the same coach and Acceleration/Pace/Agility is trained by the same coach too.
Another problem is that we don't know all good attributes. Stamina e.g. is bad according to fm-arena test results but good according to ykykyk.
Han106 said: But cheese pizza is boring and we still have 4-7 sessions that we can fill. Expand
Max said that we don't need to fill the empty spaces and he is right. I had worse training results when I used his training schedule and filled wednesday. That's why I'm looking forward to your results. Match practice performed quite well in my tests.
Yarema said: I doubt it's that simple and linear. I haven't seen any specific numbers on it, hard to quantify anyway. Expand
One source:
Spoiler
Consistency - How consistently he performs from match to match. On the basis that there is no player that will perform to his current ability 100% of the time, the ratio of what consistency means in terms of how often a player plays at his CA is as follows
20 Consistency = 20/25 games played at CA
10 Consistency = 10/25 games played at CA
1 Consistency = 1/25 games played at CA
The way consistency is used in the full match engine is that the average "off day" will be 10 below the CA. A random factor will be used to determine how much this is less or more than 10.
Please note that physical attributes are not taken into consideration when it comes to consistency. Only the technical and mental attributes are marked down depending on consistency.
Also Match Importance (database value) and Team Blend (Game value) affect the match by match calculation of the "Real CA" (Consistency combined with CA)
Don't forget Dribbling. Jumping Reach (71.6 points) and Dribbling (69.8 points) are pretty close together. And then comes Balance (65.3) which isn't negligible. And then comes Concentration (64.5
points) and then Anticipation (64.3 points) and so on.
https://fm-arena.com/table/26-player-attributes-testing/
Do you see the point? I can create a player which has e. g. lower Jumping Reach but higher Dribbling, Balance and Concentration and is therefore a better player. It is good that we have found Meta-Attributes but we should see these attributes in the right context and not in a "I don't care all other attributes" attitude because this is not adequate. Winning a league with only a few player attributes is spectacular, but it is not the optimal. Winning a league against AI managers isn't difficult. I can do this with any average team when I use the right tactic and improve team cohesion. That can't be the benchmark.
But the testing teams have only attackers with small jumping reach, haven't they?
There are a) changes and b) never forget: There isn't a minimum RNG. It is always a confidence interval like "Minimum score with 90% confidence or 95% confidence etc." so you can get a higher or lower score outside the confidence interval.
It's the power of RNG. 2,400 matches give enough room for such increases.
There was a new update yesterday. We have 24.3 now.
Then you forego the increase in stamina at half time. If you assume that the starting eleven are better than the bench players, then you should change them after half-time.
I've been testing a lot lately and have regained the joy of FM (I haven't bought an FM every year in the last few years because it's become too easy for me). But the results of tests should always be presented objectively and the Reddit user didn't do that. All attributes matter and it is interesting to find out how important each attribute is for each position. I would beat e.g. this reddit users team with my 19 acceleration, 19 pace, 19 anticipation, 19 jumping reach, 20 finishing, 20 technique, 20 first touch, 20 heading striker. This problem is more complex than reddit users posts. Yes, we found meta-attributes but this isn't the end of the research.
We can isolate single variables in FM. Any attribute is a variable. Luckily, FM is just a game and there is no reason to assume that a higher attribute value leads to worse results (regarding your glue example). When I read hints of the developers I always read things like this: "A goalkeeper's ability to save penalties is initially governed by his Anticipation, Reflexes, and Concentration in reacting to the moment the ball is struck. His Acceleration will help his immediate chances of reaching the ball successfully, whilst his Agility, Reflexes and Handling will ultimately determine whether or not he pulls off the save." So a higher value is always better and maybe all these attributes are additively linked like e.g. "0.4*anticipation+0.8*reflexes+0.3*concentration+0,2*acceleration+0.4*agility+0,3*handling" when there are penalties. We shouldn't forget that this is a game designed for entertainment. But these are all theories and assumptions. Back to practice:
I can beat machine learning with our single variable approach and I can prove it:
Spoiler I will make a team which is worse than the opponent team according ykykyk rating. ykykyk is the result of machine learning and ykykyk is based on 40 million matches. I will even use the tactic which the machine used for the 40 million matches (ZaZ - Blue 3.0). My team has only one information source: A few thousand matches on fm-arena tables which lead to the conclusion that pace and acceleration are the most important variables. So it is me against the machine:
Both teams have this starting players, bench players and tactic:
My Team is "Arsenal 23/24" and my opponent is "Arsenal 23/24 (2)". ykykyk ratings are higher for all "Arsenal 23/24 (2)" counterparts on the field and on the bench (except goalkeepers):
So "Arsenal 23/24 (2)" should beat my team according machine learning. These are the results after 40 matches:
"Arsenal 23/24" : "Arsenal 23/24 (2)"
This is a clear result beyond RNG. It isn't surprising when you consider how strong pace and acceleration are on the fm-tables. 40 million matches aren't enough for a machine to see this. This is the big advantage when you just change one variable to see what are the really important variables. Machine learning is a great tool when you are Google and have the ressources to simulate billion of billions of matches. We don't have these ressources so our approach is better than machine learning with 40 million matches.
I agree with you.
MeanOnSunday said: The only way to know at a single position level is vary multiple attributes in the same experiment and vary by position as well (this is in effect what the machine learning data is reporting).
I don't agree. You have to vary one attribute for one position to know its value for this position (in a particular tactic). If you vary multiple attributes than you don't know which attribute is important for the performance increase.
Spoiler Why do scientists change only one variable in a controlled experiment?
In a controlled experiment, scientists change only one variable at a time in order to accurately determine the effect of that specific variable on the outcome. By isolating and manipulating one variable while keeping all others constant, scientists can establish a cause-and-effect relationship between the variable they are testing and the results they observe. This helps in drawing accurate conclusions and making reliable predictions based on the experiment's findings. If multiple variables were changed at the same time, it would be difficult to attribute any observed effects to a specific variable, leading to unreliable or inconclusive results.
https://www.quora.com/Why-do-scientists-change-only-one-variable-in-a-controlled-experiment
Decision is for every position more expensive than Anticipation (except SC) but the results on the tables are worse in comparision to Anticipation (FM22/FM23/FM24). Maybe there is one position (or even two positions) where Decision has a better cost-benefit ratio than Anticipation. But that doesn't matter. Why? Because you can't train Decision just for one position. You always must train Decision for all positions:
So a high value for Decision in a training session is always a bad thing.
Individual training has no effect on attributes within a position (no difference between e.g. no-nonsense centre-back and ball playing defender). It has an effect when you train a new position e.g. right defender learns wing-back position. But this is valid for all training sessions.
My test:
Spoiler DL is trained as DL (Wing-back attack)
DR is trained as WB (Wing-back attack)
Only match practice training
Starting point of all attributes = 50 (ingame = 10)
Players are 20 years old
4 runs
It seems that DR doesn't train natural fitness and strength.
There is a big problem. One example: Anticipation is good but Decision is bad in context of CA efficiency according to the test results on fm-arena. But there is no training session which trains more Anticipation but less Decision.
Excerpt from MaxEBFM results:
Another example is Acceleration/Pace (both are good) and Agility (bad). I tried to beat the system by using a good and a bad coach but that doesn't work either. Anticipation/Decision is trained by the same coach and Acceleration/Pace/Agility is trained by the same coach too.
Another problem is that we don't know all good attributes. Stamina e.g. is bad according to fm-arena test results but good according to ykykyk.
Max said that we don't need to fill the empty spaces and he is right. I had worse training results when I used his training schedule and filled wednesday. That's why I'm looking forward to your results. Match practice performed quite well in my tests.
One source:
Spoiler Consistency - How consistently he performs from match to match. On the basis that there is no player that will perform to his current ability 100% of the time, the ratio of what consistency means in terms of how often a player plays at his CA is as follows
20 Consistency = 20/25 games played at CA
10 Consistency = 10/25 games played at CA
1 Consistency = 1/25 games played at CA
The way consistency is used in the full match engine is that the average "off day" will be 10 below the CA. A random factor will be used to determine how much this is less or more than 10.
Please note that physical attributes are not taken into consideration when it comes to consistency. Only the technical and mental attributes are marked down depending on consistency.
Also Match Importance (database value) and Team Blend (Game value) affect the match by match calculation of the "Real CA" (Consistency combined with CA)
https://fmmvibe.com/forums/topic/8738-player-hidden-attributes-explained/