It would be interesting to see how much the other attributes matter at lower thresholds. For instance, all the sample players have 13 decisions by default (pretty good score), so +- 5 ranges from 8 - 18. It might very well be that 8 decisions is "sufficient" and there would be a large dropoff at even lower values. Something like 5 - 15 could be cool to see for many of these. Especially, it wouldn't surprise me if the difference between 13 and 18 is rather miniscule for many of these attributes, as 13 is a very good number already.
I'm super fascinated by the ykykykyky findings and although my conclusion remains that they are very role/instruction specific for ZaZ Blue, I have no doubts that they are as near optimal as you can get for that exact application, and it should be usable for other tactics with overlapping roles/instructions.
Some of the weights surprise me though, take the Striker position for example: - Agility is often touted as a very important attribute (and appears to be) for nearly every position, yet the AF(at) has it weighted at only 30? Keep in mind that the official CA weights have it at 6.0/10.0 and the role even has it marked as key. - Technique is weighted at a whooping 65, although it is an attribute that scores poorly on the attribute test here on fm-arena. Hell, it's weighted around 50 at an average across all positions in ykykyky, which really makes me wonder why it doesn't have more impact on the attribute tests? My only plausible explanation at this moment is that there must be some widely used player/team instruction in ZaZ Blue that places a huge emphasis on Technique which the attribute testing tactic does not have... I see the former has slightly more direct passing, could that be it?
More generally, I have some thoughts about other attributes as well: - Aggression and Work Rate is rated high across every position, but this makes sense due to ZaZ blue being having super high press, high tempo and has harder tackles enabled. These results show us that these attributes need to be considered when employing such tactics (which are super popular in FM). - Decisions is generally weighted almost half of what the CA Attribute Weights say, even though all positions in the tactic have Take More Risks enabled. I believe this proves that decisions isn't nearly as important as the game wants us to believe, even when the players are given creative freedom. A natural attribute to "cut" for better CA spread. - Physicals are king, but we already knew that. - Flair isn't very important, even if key to a role and much creative freedom is allowed. - OTB seems to be very important for roles that cut inside (IWB's in ZaZ). I suspect AMR/L's who are deployed as Inside Forwards or IW's would have a greater reliance on this attribute as well. Even so, it is generally a good stat to have on all players, perhaps because it combines well with superior physicals? - Determination has some effect, but not as much as one might be lead to believe. An okay increase but nothing to prioritize a lot (might be because the players develop better across a season?) - Teamwork scores relatively well compared to what I believed before, maybe due to high press? - Bravery seems to be the least important DNA attribute, and ZaZ even uses harder tackles which should increase its value. Not worth considering?
It's hard to make any definite statements about the weights, but they have definitely changed the way I view the game. At least it proves that there are many attributes that could be considered "free" for a position that have the potential to make a huge difference. Player DNA really is a thing, especially Work Rate and Aggression.
I used the ykykyky balanced method religiously for some seasons as Leicester albeit with a different tactic (still 451). I have no doubt the ML findings are very accurate for zaz blue, and that they still work fine for other setups as well, but I do think that it is highly probable to not be as efficient for tactics using other roles, duties, instructions and/or formations. ML by nature will only optimise upon the parameters given, and if millions of games were simulated using only the same tactic (which I believe was the case here?), instructions cease to matter as a variable factor to the system and the results.
This leaves us with two possibilities, either instructions relative to attributes doesn't matter in the engine (unlikely) or ykykyky represents the optimal weights for its given set of instructions alone. This makes sense as you can't expect a target man striker in a 442 to be dependent on the same attributes as an advanced forward in a 451. If this was the case, the game engine would be a joke.
However, there is a lot of overlap with the positional weights given by the editor, as well as the fmarena attribute tests, so at least it somewhat proves that core attributes such as acc and pace truly are as significant as claimed. There is also without doubt value in looking closer at the attributes that differ greatly from these other weightings, but then specifically in conjunction with the team and player instructions used in zaz blue.
So, if you're using zaz blue, these reported attribute weightings truly must be as close to optimized as you could get. But I'm not convinced that it translates strictly better than the other attribute weight-systems on a general basis... If anything, I would maybe rely more on @Marks approach with having a cutoff and only consider the very best attributes reported per position (maybe at 50+, or even higher) rather than using the whole GS rating spread from 0-100 as the lower you go the more likely it is that an attribute is (significantly) affected by specific instructions, or not dominant enough to be considered too much with the uncertainties we are operating with for a generalized strategy.
Let me just re-iterate that I don't mind being wrong on this, but my findings from exclusively collecting the "best" players in every position as reported by GS using ykykyky didn't consistently improve results or ratings in my differing system. This is an anecdotal observation of course, as I've run no systematic test to prove or disprove ykykyky balanced as the best generalized weighting choice.
I do believe the scientific approach is incredibly cool, and I hope that this isn't the last large scale machine learning optimization project we see on the matter, as the use case for such methodologies is incredibly suitable for a game such as FM. I imagine that running a similar test, but using a set of (vetted) tactics that all perform relatively well with differing roles and instructions to find the true averages of attribute-importance could be mind-blowing.
Edit: As an addendum, I will probably be running a LLM game and implementing the @Mark approach to low-CA-cost-to-performance filters that he suggested as I think this is a very good use case for these findings. I'll let you know how it pans out.
@Mark Hey, really nice work and I'm sorry if I'm asking a stupid question but with all the different rating systems I'm still not sure how to actually select attributes for my filters in-game. I have genie scout with ykykyky ratings but I don't want to rely on that while playing.
I see that the attributes are weighted differently, but how many of the top attributes do I filter on? For instance, GK has Agi 100, Reflexes 80, Acc 70, Strength 70, Concentration 65, Aerial 60, Decisions, Pace and Handling 50. I suppose taking all of these is too many, so where should I cut? Above 50? And then for instance filter agi at avg+2, reflexes avg+1, acc avg+1, str avg+1, conc avg, aerial avg? I'm generally playing in top divisions so I'm only looking for the best of the best.
I'd really appreciate some input on the actual process here, like how you would go about building an in-game GK filter, so that I can replicate the process for the other positions myself.
Some of the weights surprise me though, take the Striker position for example:
- Agility is often touted as a very important attribute (and appears to be) for nearly every position, yet the AF(at) has it weighted at only 30? Keep in mind that the official CA weights have it at 6.0/10.0 and the role even has it marked as key.
- Technique is weighted at a whooping 65, although it is an attribute that scores poorly on the attribute test here on fm-arena. Hell, it's weighted around 50 at an average across all positions in ykykyky, which really makes me wonder why it doesn't have more impact on the attribute tests?
My only plausible explanation at this moment is that there must be some widely used player/team instruction in ZaZ Blue that places a huge emphasis on Technique which the attribute testing tactic does not have... I see the former has slightly more direct passing, could that be it?
More generally, I have some thoughts about other attributes as well:
- Aggression and Work Rate is rated high across every position, but this makes sense due to ZaZ blue being having super high press, high tempo and has harder tackles enabled. These results show us that these attributes need to be considered when employing such tactics (which are super popular in FM).
- Decisions is generally weighted almost half of what the CA Attribute Weights say, even though all positions in the tactic have Take More Risks enabled. I believe this proves that decisions isn't nearly as important as the game wants us to believe, even when the players are given creative freedom. A natural attribute to "cut" for better CA spread.
- Physicals are king, but we already knew that.
- Flair isn't very important, even if key to a role and much creative freedom is allowed.
- OTB seems to be very important for roles that cut inside (IWB's in ZaZ). I suspect AMR/L's who are deployed as Inside Forwards or IW's would have a greater reliance on this attribute as well. Even so, it is generally a good stat to have on all players, perhaps because it combines well with superior physicals?
- Determination has some effect, but not as much as one might be lead to believe. An okay increase but nothing to prioritize a lot (might be because the players develop better across a season?)
- Teamwork scores relatively well compared to what I believed before, maybe due to high press?
- Bravery seems to be the least important DNA attribute, and ZaZ even uses harder tackles which should increase its value. Not worth considering?
It's hard to make any definite statements about the weights, but they have definitely changed the way I view the game. At least it proves that there are many attributes that could be considered "free" for a position that have the potential to make a huge difference. Player DNA really is a thing, especially Work Rate and Aggression.
This leaves us with two possibilities, either instructions relative to attributes doesn't matter in the engine (unlikely) or ykykyky represents the optimal weights for its given set of instructions alone. This makes sense as you can't expect a target man striker in a 442 to be dependent on the same attributes as an advanced forward in a 451. If this was the case, the game engine would be a joke.
However, there is a lot of overlap with the positional weights given by the editor, as well as the fmarena attribute tests, so at least it somewhat proves that core attributes such as acc and pace truly are as significant as claimed. There is also without doubt value in looking closer at the attributes that differ greatly from these other weightings, but then specifically in conjunction with the team and player instructions used in zaz blue.
So, if you're using zaz blue, these reported attribute weightings truly must be as close to optimized as you could get. But I'm not convinced that it translates strictly better than the other attribute weight-systems on a general basis...
If anything, I would maybe rely more on @Marks approach with having a cutoff and only consider the very best attributes reported per position (maybe at 50+, or even higher) rather than using the whole GS rating spread from 0-100 as the lower you go the more likely it is that an attribute is (significantly) affected by specific instructions, or not dominant enough to be considered too much with the uncertainties we are operating with for a generalized strategy.
Let me just re-iterate that I don't mind being wrong on this, but my findings from exclusively collecting the "best" players in every position as reported by GS using ykykyky didn't consistently improve results or ratings in my differing system.
This is an anecdotal observation of course, as I've run no systematic test to prove or disprove ykykyky balanced as the best generalized weighting choice.
I do believe the scientific approach is incredibly cool, and I hope that this isn't the last large scale machine learning optimization project we see on the matter, as the use case for such methodologies is incredibly suitable for a game such as FM. I imagine that running a similar test, but using a set of (vetted) tactics that all perform relatively well with differing roles and instructions to find the true averages of attribute-importance could be mind-blowing.
Edit: As an addendum, I will probably be running a LLM game and implementing the @Mark approach to low-CA-cost-to-performance filters that he suggested as I think this is a very good use case for these findings. I'll let you know how it pans out.
I see that the attributes are weighted differently, but how many of the top attributes do I filter on? For instance, GK has Agi 100, Reflexes 80, Acc 70, Strength 70, Concentration 65, Aerial 60, Decisions, Pace and Handling 50. I suppose taking all of these is too many, so where should I cut? Above 50? And then for instance filter agi at avg+2, reflexes avg+1, acc avg+1, str avg+1, conc avg, aerial avg? I'm generally playing in top divisions so I'm only looking for the best of the best.
I'd really appreciate some input on the actual process here, like how you would go about building an in-game GK filter, so that I can replicate the process for the other positions myself.