Optimizing AI Squad-Building: The tactics creator offers the best way forward

Recently, I took a look at AI squad-building over a period of 10 years and came to the conclusion that, for the most part, the current system works fairly well. The FM12 AI generally improves its squad in all the necessary positions, avoids oddities with terrible flaws (for example, centrebacks that can't jump) and favors, likely as a result of the player development system, players with very good ratings for Professionalism and Determination. On the whole, particularly for big teams, a team's squad in 2022 will be better than the squad from 2011 as a result of average CA increases and overall improvements to squad personality.

With that said, there is room for improvement, and the obvious area demanding optimization is the means by which the AI selects its transfer targets.

Currently, the AI's transfer policy relies on the relatively simple CA system in which players are given a positional rating based on the broadest concept of that position. While this system generally works well assuming you're looking at a jump of 20 or 30 CA points, it is oversimplified for instances in which the AI is attempting to make finer improvements (as tends to be the case after about 8 years when clubs hit a ceiling for average CA). When presented with two players with very similar CA ratings, the AI manager currently has no means of determining which player will actually be better for his tactical system. Thus, a Tony Pulis may often go for a small though technically gifted striker for his long-ball "anti-football" approach if his CA is just slightly better than a powerful target man with terrific mental attributes, even though the latter is likely to be a much better fit for his tactical system. While there are obviously instances where this happens in real life (Ibra at Barcelona comes to mind), it seems to happen far too frequently in Football Manager. While an "Ibra mistake" would more accurately translate to the AI deciding that a 15-20 overall boost in CA outweighs stylistic preferences, the problem is that every AI transfer under the current system is basically a shot in the dark in terms of tactics.

And this, in turn, relates to the fact that the AI appears not to have any clear sense of more specific player roles. It uses the general parameters of classic tactics, and thus, it sees a mid-fielder as a mid-fielder, not as an advanced playmaker or a ball-winner or "just" a generic supporting mid-fielder.

Fortunately, SI has already created the system that the AI needs to vastly improve its squad-building behavior. While many FM players feel the tactics creator is overly simplified and prefer tweaking every individual slider to fine tune their precise vision, the classic sliders create the opposite problem for the AI. They're too general, they lack specificity and they ultimately lead to decisions based solely on the most general assessments of player quality. Moreover, as the AI doesn't fine tune individual positions as a human player does, this causes AI tactics to lack truly distinct personalities. An aggressive, short-passing 4231 is an aggressive short-passing 4231, you won't see one where a playmaker sits in the gap to feed balls to an adventurous complete forward and another where a roving trequartista works with a more conservative poacher.

However, the tactics creator and its specifically defined roles provide a means to give more individuality to each AI manager while enabling the AI to make more precise decisions about relative player quality.

So how should SI go about this?

First, AI managers should no longer be defined by "tactical attributes" corresponding to universal sliders. Rather, the managers in the initial database should have a primary tactic defined by the same formation, team instructions and assigned player roles that players use in the tactics creator with further flexibility provided by "offensive" and "defensive" options as well as mid-match strategy changes and touchline shouts. For staff without assigned tactics and newgen staff, their tactical preferences should not be randomly generated based on a single, generic national template but drawn from an expandable (via the official editor) list of preset tactics assigned to each nation. This would have the additional effect of enabling more variability among newgen tactics and also avoid the often incoherent and grossly ineffective tactical arrangements that result from the random assignment of tactical tendencies.

Second, AI managers should no longer merely look at reputation and CA/PA when making line-up and transfer decisions. Rather, it should build a short list based on specific positional needs and a broad CA range, then make a more specific decision based on a secondary CA rating that accounts solely for the "important attributes" of the role it is looking to fill. This secondary rating should be balanced with the more general assessment to ensure that the AI will often go for slightly lower-CA players better suited to its tactic while still occasionally opting for a bigger star who doesn't necessarily fit the system perfectly (again, the Ibra example). Additionally, players who lack "Accomplished" or higher positional competence in any position used in the manager's primary tactic should be excluded from consideration altogether.

Third, the more "flexible" AI managers should be equipped to recognize when all the available options for a role are grossly inadequate (for example, when a slow CM who can't cross is the only healthy option for an MR position set to "Winger - Attack" or a pacey young striker is the only healthy option for a ST set to "Target Man") and temporarily (i.e., at the beginning of a match) alter the role/duty to something better suited to the player. And if the manager lacks an adequate option for the role altogether (due either to the absence of any such player or 6+ month injuries to a preferred starting player), filling that role with a suitable replacement should be flagged as a transfer priority.

Finally, though somewhat tangential, AI managers must be equipped to take into account homegrown rules for all competitions for which their club may qualify. Thus, managers whose teams are well below possible homegrown standards should exhibit an extreme bias towards signing the best viable domestic players while increasing the acceptable transfer fee for any existing homegrown talent.

While this would obviously constitute a huge evolution for the game, it is the next necessary step that must be taken to address concerns about the difficulty of long-term saves. While recent adjustments to morale have made things more difficult in the short-term, this challenge quickly dissipates after several successful seasons when a player's reputation has essentially maxed out and AI squad-building behavior hits its CA ceiling. To truly create a consistent and significant challenge, the AI must learn how to identify and build towards a coherent tactical vision.

I don't have time to give a long comment, but I think it's great that things like these are discussed. These issues that you adress are what, in my opinion, are currently holding the game back the most, because the it turns predictable and unrealistic after a while.

It'd be nice with a separate forum for development ideas where people who really cared could discuss, and perhaps even get some feedback from the developers, when they found something interesting.

Best post I've seen, people always want ******** improvements but this is actually the biggest thing wrong with FM. This needs to be priority, it's not gonna change massively from one year to the next but by lets say the 2015 edition there should be drastic changes to the way AI works

Very good post there The Hand of God:thup:

I agree that some way of making sure managers buy players suited to their tactics and play them in a way that gets the most out of them is needed. The Tactics creator and it's player roles would seem like a sensible way to implement that, And id be delighted if it turned up in FM13. But what Id like to see more long term would be AI managers adapting their preferred formations to incorporate out standing talents. As it stands now if an AI manager has three preferred formations none of which utilise a AM and he has the worlds best AM he will either play him out of position or leave him out entirely it would be nice if preferred formation could be added and removed depending on who the manager has at his disposal.

Terrific post OP. I wholly agree.

Completely agree with you here. I have long requested that a player-type system connected to both tactics and player scouting is implemented into the game. If a tactic has a deep-lying defensive midfielder playmaker, its manager should be able to search for such a player in the scouted database. This should replace the simple "flamboyant midfielder" descriptions. The AI would benefit greatly from this, as they would be making more informed decisions regarding both signings and team selection.

would be nice if someone from SI would take notice of this post.

Hey..

Don't understand this point: generic national template but drawn from an expandable (via the official editor) list of preset tactics assigned to each nation.

does that mean that they take the preset from the nations settings?

Just on the tactic issue, PaulC said in another thread that the AI doesn't use the Tactics Creator in way we do because it wouldn't work for it (perhaps exactly for the reasons you mention).

And whilst your suggestions make sense, I think it is way beyond the scope of the game in the immediate future. The game pans-out the way it does currently because in order to get a game world functioning the AI has to use a pool of patterns and routines. To want to have AI managers looking at each team's strengths and weaknesses, *then* to make decisions based on that analysis (and to have that analysis differ to allow for 'better' managers to exist) and have such decisions reviewed for each game based on opposition and for transfer targets (to make the tactics element worthwhile) would take a massive amount of processing power and coding routines.

Hey..

Don't understand this point: generic national template but drawn from an expandable (via the official editor) list of preset tactics assigned to each nation.

does that mean that they take the preset from the nations settings?

Currently, regen staff tend to take their nation's tactical settings and then slightly randomize each attribute. Instead of randomized tactics (many of which just don't work), I would prefer regen staff draw from an expandable pool of proven, tested tactics based on historic tactics, current managers and user-made tactics (imagine facing a new manager using a super-tactic from these forums, for example).

And whilst your suggestions make sense, I think it is way beyond the scope of the game in the immediate future. The game pans-out the way it does currently because in order to get a game world functioning the AI has to use a pool of patterns and routines. To want to have AI managers looking at each team's strengths and weaknesses, *then* to make decisions based on that analysis (and to have that analysis differ to allow for 'better' managers to exist) and have such decisions reviewed for each game based on opposition and for transfer targets (to make the tactics element worthwhile) would take a massive amount of processing power and coding routines.

I wouldn't expect to see anything like this in 12.3, but in the next few years, I think most people hope to see some major improvements taken in the transfer area.

As for the processing power issue, I don't think this approach would differ too much from what already happens. The AI already does analyze its positional weaknesses, it already creates shortlists for active AI managers and it even analyzes the individual tactical weaknesses of the opposition. To make the above approach work, you will need to add the following behaviors:

(1) CA-based rating of a player based on each role, so basically, an additional set of attributes for each player, initially generated at the beginning of each game and when regens are produced

(2) Exclusion of transfer targets that don't play a position in an AI manager's preferred formation

(3) Recognition when team falls below a certain "homegrown" threshold and appropriate adjustment to the existing "Domestic Player Bias" AI tendency

(4) Recognition when a role-rating for available players falls below a certain relative threshold and adjustment to a role in which a player has a higher rating (and this would only need to occur in the one or two leagues that are set to full detail)

These aren't really complex behaviors, just a few additional rules and one subsystem based on the existing system for analyzing players.

In regards to there being "better" managers, as I understand it, transfer behavior is the same for all managers and I wouldn't have a problem with that remaining the case. Where AI managers differ in skill is their man-management behavior and that appears to be functioning very well.

And given how much processors have advanced since the current system was designed, I don't think this would create too great a burden. I played FM11 with multiple loaded leagues and 80,000+ players on a 6 year-old laptop and didn't have any issues running the game. On my new laptop, FM12 is extremely fast, even with nearly all leagues loaded and most international competitions set to full detail.

Bit of a big bump here... But before I go out and buy FM2015, can anyone tell me if the AI squad building has been improved in any way? The last FM I played was FM2011 or FM2012 (not sure anymore) and I remember that getting boring after a while with AI teams becoming too weak to offer a good challenge.

I'm hoping that won't be as much of a problem if I start unemployed with low reputation and go for a journeyman save, which is the plan, but I'd still like to hear some opinions on if and how AI squad building has been improved in recent editions of FM and especially in FM2015.

A think this thread needs to be looked at again, with the big problem right now being AI Squad building, We need to take this into account and promote a discussion on this a such.

AI squad building is no different this year sadly