Currently existing levers for managing AI difficulty and their values

[oleg-derevenetz]

Related to #8270
Updated after #8402

In this topic, I provide a complete list of currently available levers for managing AI difficulty and their current values. @zenseii @Branikolog @LeHerosInconnu please feel free to discuss these levers and difficulty level hierarchy (i.e. their aims, target audiences, etc) with your proposals of specific values for these bonuses (or limitations) for each difficulty level here.

• Allow AI to split weak stacks by 1: Easy & Normal - NO, other difficulties - YES;
• The ratio of the strength of the enemy army to the strength of the AI army, above which the AI decides to surrender or retreat from the battlefield: Easy - 100 / 6, Normal - 100 / 7.5, Hard & Expert - 100 / 8.5, Impossible - 100 / 10;
• Whether it is allowed for AI to build specific castle buildings:
  • Campaign: always YES;
  • Regular maps: Easy - NO for level 6 dwellings and level 5 mage guild, otherwise YES, other difficulties - always YES;
• Whether it is allowed for AI to build castle buildings depending on day:
  • Campaign: always YES;
  • Regular maps: Easy - allowed only on even days, other difficulties - always YES;
• Whether it is allowed for AI to assign roles to heroes: Easy - NO, other difficulties - YES;
• Minimum advantage in stats (i.e. the sum of the levels of primary and secondary skills) that a hero must have in order to allow another hero with the same role to meet on his own initiative with this hero to exchange armies and artifacts: Easy - 10, other difficulties - 2;
• Bonus resource income for AI-controlled kingdoms:
  • Easy & Normal - no bonus;
  • Hard and Expert - +1000 gold per day;
  • Impossible - +2000 gold per day;
  • Additionally, on Expert & Impossible difficulties, AI will get +1000 gold per castle regardless of its type and dwellings. Additionally on these difficulties, if a corresponding castle has a level 6 dwelling, the following per-castle resource bonuses may be applied:
    • +1 Crystal per day for each Barbarian castle - only if the kingdom has a Marketplace in at least one of its castles or has at least one captured Crystal mine;
    • +1 Mercury per day for each Sorceress castle - only if the kingdom has a Marketplace in at least one of its castles or has at least one captured Mercury mine;
    • +1 Sulfur (only if the kingdom has a Marketplace in at least one of its castles or has at least one captured Sulfur mine) & +1000 Gold per day for each Warlock castle;
    • +1 Gem (only if the kingdom has a Marketplace in at least one of its castles or has at least one captured Gems mine) & +1000 Gold per day for each Wizard castle;
• Bonus gold income for AI-controlled kingdoms (applied after applying the resource income bonus): Easy - -25% (i.e. deduction from the total income), other difficulties - no bonus;
• Limit on the number of Dimension Door spells that AI-controlled hero can cast during the single turn: Easy - 1, Normal - 2, Hard - 3, other difficulties - unlimited.

[Mr-Bajs]

11. Scouting bonus for AI-controlled heroes: Easy - no bonus, Normal - +1, Hard - +2, Expert - +3, Impossible - +4;

I think the general perception on Normal difficulty in games are that the AI more or less have the same or have the goal to be as equal to the player in terms of buffs and nerf. Shouldn't then the normal difficulty also have a no bonus, and then maybe Hard lowered to +1 and Expert to +2 and Impossible +3?

5.2. Regular maps: Easy - allowed only on even days, other difficulties - always YES;
I suppose this refers to weekdays? Then the AI can't build on the 7:th day an not the day after that's the 1:st. Two days in a row where the AI can't build is that intended? That that means it can only build on 2, 4, 6 week days, can only build on 43% of the days. Isn't that maybe to severe?

6. Whether it is allowed for AI to assign roles to heroes: Easy - NO, other difficulties - YES;
Could that lead to some unwanted weird behavior?

[oleg-derevenetz]

I think the general perception on Normal difficulty in games are that the AI more or less have the same or have the goal to be as equal to the player in terms of buffs and nerf.

This is exactly what is proposed to be discussed here :)

I suppose this refers to weekdays?

No, this refers to "end-to-end" numbering of days starting from 1. So 1-2-1 is day 8, 1-2-2 is day 9, etc.

Could that lead to some unwanted weird behavior?

This will lead to the fact that the map scouting and reinforcement of the heroes with troops brought from the castles will work worse than with roles. But that's the point of easier difficulties, right?

[zenseii]

Hi, @Mr-Bajs.

Two days in a row where the AI can't build is that intended? That that means it can only build on 2, 4, 6 week days, can only build on 43% of the days. Isn't that maybe to severe?

From my tests it isn't too severe, and in my opinion it is better to talk about these changes in terms of testing them in real playtests to see how they behave, rather than on a general principle of the change. You should approach easy difficulty as a difficulty aimed for complete novices to the game, and they will inevitably not construct on some days or construct completely "wrong" things, like spend a day to construct a farm which is essentially the same as wasting a day of construction and wasting 1000 gold.

[LeHerosInconnu]

Hello everyone,

So, here are my first impressions for this:

1. Allow AI to split weak stacks by 1: Easy & Normal - NO, other difficulties - YES;

Making troops of one creature is not necessarily a good thing, as it can give the opponent the hand the following turn, which can be devastating.
As a general rule, it's a tactic that only really makes sense if you're planning to surrender or flee.
It's also, in my opinion, better suited to defense than attack.
For a powerful attacker, for example, it's better to have three troops of 100 goblins rather than one troop of 298 goblins and two troops of 1 goblin, as this gives more tactical possibilities.
The trick of absorbing the retaliation of an opposing troop with a troop of 1 creature is a little weak.
Of course it works, but in the end, it's not really that important.
Heroes 2 is not Heroes 3, there's no possibility of making a troop wait, flying troops cross the entire battlefield, ranged shooters do their damage regardless of distance, the difference between attack (from attacker) and defense (from defender) is more important, etc.
In Heroes 2, it's better to have big troops that hit hard and hit first.

2. The ratio of the strength of the enemy army to the strength of the AI army, above which the AI decides to surrender or retreat from the battlefield: Easy - 100 / 6, Normal - 100 / 7.5, Hard & Expert - 100 / 8.5, Impossible - 100 / 10;
3. The minimum ratio of the AI kingdom's gold reserve to the cost of surrender, at which the AI will prefer surrender to retreat from the battlefield: always 10;

15. The minimum hero level at which the AI can consider the possibility of surrender or retreat from the battlefield for this hero: always 3.

The hero should flee (or surrender) when he considers that he can no longer do damage to the opponent without running the risk of no longer being able to flee.
In short, if he has no chance of decimating at least one opposing creature without risking not being able to flee, then he flees.

The action of surrendering should rather be evaluated according to the creatures already recruited by the player (with his other heroes and garrisons), the creatures available for recruitment by the player (in his cities) and the level of gold generated by the player (each day).
So, in the majority of cases, only level six creatures (perhaps lower level creatures at the start of the scenario when level six creatures are not yet available) are really worth keeping, while other creatures are like gold given to the opponent (as soon as level six creatures are available).
There's also the special case of the strong hero with good artifacts defending a castle, where it's worth surrendering even with lower level creatures.
The aim of surrender should be to conserve high level troops if they are not easily accessible, and to conserve important artifacts or a high hero in the case of city defense.

Per example, if the number of creatures in the army of the hero who could surrender is less than 30% (can be adjusted, maybe 25% or 20%) of the number of creatures the player owns with his other heroes and garrisons, added to the number of creatures available for recruitment (thus excluding those in the army of the hero who could surrender), and if the player has an amount of gold plus a daily income that allows him to own and/or recruit at least 50% (can be adjusted) of the creatures before the end of the current week, then there's no need for the hero to surrender (he can simply flee).
In my opinion, there should be no difference for all levels of difficulty.

4. Whether it is allowed for AI to build specific castle buildings:
   4.1. Campaign: always YES;

I am in favor of no restriction for campaigns.

4.2. Regular maps: Easy - NO for level 6 dwellings, otherwise YES, other difficulties - always YES;

In easy mode only:

• AI does not build the dwellings of level six creatures.
• AI does not build level five mage guilds.

5. Whether it is allowed for AI to build castle buildings depending on day:
   5.1. Campaign: always YES;
   5.2. Regular maps: Easy - allowed only on even days, other difficulties - always YES;

I don't think there's any need for that.
Prohibiting the AI from building the dwelling of level six creatures and level five mage guilds is enough.

Note: AI can still recruit level six creatures in dwellings built by the human player and can learn level five spells in mage guilds built by the human player.

6. Whether it is allowed for AI to assign roles to heroes: Easy - NO, other difficulties - YES;

Okay, but not necessary with prohibiting the AI from building the dwelling of level six creatures and level five mage guilds.

7. Minimum advantage in stats (i.e. the sum of the levels of primary and secondary skills) that a hero must have in order to allow another hero with the same role to meet on his own initiative with this hero to exchange armies and artifacts: Easy - 10, other difficulties - 2;

Same, not necessary with prohibiting the AI from building the dwelling of level six creatures and level five mage guilds.

8. Bonus growth in castle dwellings controlled by AI: Easy & Normal - no bonus, Hard - +14%, Expert - +25.4%, Impossible - +36.8%;

I thought this had been completely removed.
It's more a benefit to the human player than anything else.

9. Movement points bonus for AI heroes: Expert & Impossible - +75MP, other difficulties - no bonus;

11. Scouting bonus for AI-controlled heroes: Easy - no bonus, Normal - +1, Hard - +2, Expert - +3, Impossible - +4;
12. Limit on the number of Dimension Door spells that AI-controlled hero can cast during the single turn: Easy - 1, Normal - 2, Hard - 3, other difficulties - unlimited.

Personally, I would remove these bonuses and limitation.
But I could see some bonuses and guidelines depending on the "personality" of the AI players (builder, warrior, explorer).
More on this later (#8817).

10. Bonus gold income for AI-controlled kingdoms: Easy - -25% (i.e. deduction from the total income), Normal - no bonus, Hard - +29%, Expert - +45%, Impossible - +60%;

I would do it this way:

• Easy: -25% gold income
• Normal: no bonus/malus
• Hard: +50% gold income
• Expert: +100% gold income, +1 of each resource per day
• Impossible: +200% gold income, +1 of each resource per day

13. Allow AI to retreat: always YES;
14. Allow AI to surrender: always YES.

I am okay with that (except that I'd like the human player to be able to decide whether or not to accept the AI's surrender (which could still flee if the human player refuses the surrender (except in defense during a siege)).

A suggestion concerning the use of the marketplace:

• In easy mode, AI does not use the marketplace in its cities (this may be still possible at trading post on the adventure map).
• In normal and hard modes, AI only uses the marketplace in one of its cities if it has at least 3 marketplaces, unless this allows it to build a high level building, such as a level six dwelling or a level five mage guild, this on the last day of the week.
• In expert and impossible modes, AI can use marketplace as soon as it owns at least one, provided it has significant quantities of resources and gives priority to gold for resource exchange.

[oleg-derevenetz]

I'm not a programmer

Probably that's why everything doesn't seem complicated for you :)

So I'm going to let someone who is interested in the problem and who might be able to solve it to do it for his own pleasure.

So am I. Words are cheap, let someone show the working code.

If these bonuses have come out of nowhere, I suggest removing them for the time being and seeing how things go in the game without them, so as to start from a known, understandable base.

I can do it in the scope of some PR intended for testing purposes, but I'm not sure that we should do it in the master branch right away. @ihhub @zenseii what do you think?

For evaluation purposes, would it be possible to generate a text file indicating the operations performed by the AI when using the marketplace, even if only in debug mode?

Yes, of course, it's all possible in debug builds.

Remember, I'm not a programmer, so if someone is interested in the problem and might be able to solve it, free for him to do it for his own pleasure.

Let's make it clear: if someone wants to improve somehow the marketplace usage by AI, they are free to do it. If there are specific obvious recurring problems with the current mechanism of using marketplaces, I can look at it. But if someone wants to make some improvements, then they'd better sit down and make an effort to figure out the current code and how it works, make tests, etc.

[LeHerosInconnu]

@oleg-derevenetz,

I'm not a programmer

Probably that's why everything doesn't seem complicated for you :)

So I'm going to let someone who is interested in the problem and who might be able to solve it to do it for his own pleasure.

So am I. Words are cheap, let someone show the working code.

If these bonuses have come out of nowhere, I suggest removing them for the time being and seeing how things go in the game without them, so as to start from a known, understandable base.

I can do it in the scope of some PR intended for testing purposes, but I'm not sure that we should do it in the master branch right away.

It's a good start, so I'll be able to run a few tests.

For evaluation purposes, would it be possible to generate a text file indicating the operations performed by the AI when using the marketplace, even if only in debug mode?

Yes, of course, it's all possible in debug builds.

Being able to see what AI does with it is already a good start.
Fortunately, it doesn't seem too complicated to do. Am I right?

Remember, I'm not a programmer, so if someone is interested in the problem and might be able to solve it, free for him to do it for his own pleasure.

Let's make it clear: if someone wants to improve somehow the marketplace usage by AI, they are free to do it. If there are specific recurring problems with the current mechanism of using marketplaces, I can look at it. But if someone wants to make some improvements, then they'd better sit down and make an effort to figure out the current code and how it works, make tests, etc.

In the example you've shown above, there isn't really a clear advantage or disadvantage to AI's use of the marketplace.
I won't do any tests on the subject, time is not cheap.

[oleg-derevenetz]

It's a good start, so I'll be able to run a few tests.

OK, I'll create the corresponding PR without any bonuses related to unit growth or income later today.

Fortunately, it doesn't seem too complicated to do. Am I right?

Yeah, just get the sources and make the debug build.

I won't do any tests on the subject, time is not cheap.

Well, then that's it. When I make a suggestion to improve something, I provide ready-made code that can be tested, with an explanation and demonstration (if possible) of the specific problems that it solves, and my time is not cheap either.

[LeHerosInconnu]

@oleg-derevenetz,

It's a good start, so I'll be able to run a few tests.

OK, I'll create the corresponding PR without any bonuses related to unit growth or income later today.

Fine.

Fortunately, it doesn't seem too complicated to do. Am I right?

Yeah, just get the sources and make the debug build.

I won't do any tests on the subject, time is not cheap.

Well, then that's it. When I make a suggestion to improve something, I provide ready-made code that can be tested, with an explanation and demonstration (if possible) of the specific problems that it solves, and my time is not cheap either.

It's not possible in game to access what the AI is doing with its resources or marketplace.
Apart from creating a very specific scenario using thieves' guilds as the only estimation tool, to arrive in any case at a result that will prove nothing in itself and would take a considerable amount of time.
So why test it under these conditions.

[oleg-derevenetz]

If these bonuses have come out of nowhere, I suggest removing them for the time being and seeing how things go in the game without them, so as to start from a known, understandable base.

#8402 removes all the unit growth bonuses and income bonuses for AI for all difficulty levels (all these bonuses are set to zero).

[zenseii]

I'm posting here (with permission) a discussion I had with @Stisen1 in our Discord server today which I think is very relevant:

Stisen:
My 6 year old son is getting his ass kicked before he even get a basic understanding of the game. Even on easy! Og course. Who can blame him... He's only 6! BUT... Are there any maps that have the human player start with an upper hand? He wants to steamroll a map, in order to become motivated to learn how to actually play strategically and clever. He is a clever kid, after all

Zenseii:
Someone did post a list of maps they recommend for newcomers so I'll try to look that up. Could perhaps Skrimish on easy be manageable? Maybe it's too small and boring?
Then there's Dungeon Below which pretty much let's you play at your own pace. We're currently discussing how to adjust the AI for the various difficulties, and @LeHerosInconnu is suggesting that we remove many of the limitations on the AI even on easy.

Stisen:
He did do that map. It took him 9 days to beat on easy 🙂 He'd like some more exploration

Zenseii:
Your input on this matter would be extremely valuable @Stisen11 since you have first-hand experience

Stisen:
First hand experience?

Zenseii:
Yeah that's what I figured. It's too short to be much fun. So maybe Dungeon Below? Still quite small though. I guess one advice would be to just edit one of the maps and give your kingdom lots of gold and advantages.

Zense:
-Reply to: First hand experience?
Yes, you have seen yourself what newcomers can struggle with when playing the game.
Meanwhile most of us are just basing those adjustments on assumptions or our own rather subjective experience with the game.
Here is the ongoing discussion:
#8321
Your feedback, especially regarding removal of AI limitations on easy mode, would be very valuable.

The discussion about higher difficulties than easy isn't as relevant at the moment.

Stisen:
Yeah... The AI player is way to smart for my son to handle. He's able to win battles with 50% loss of troops, but he looses another 25% or so, when he play the fights himself.

I think his intuition is that manageable AI heroes would try to flee from him, but as they move closer and closer, he doesn't trust his reasoning. You know, like when playing chess. If an easy AI opponent moves his queen into an obvious dangerous spot, an inexperienced player would think that the AI did a smart move, because they expect a hard opponent due to the massive jump between difficulty settings.

In the original Heroes 2 I remember being way more comfortable learning how to best the AI heroes. Because they acted the same way all the time, regardless of how strong their forces were. My son gets frustrated, when he sees that the AI behavior changes all the time. He'd much rather have an opponent that acts consistently stupid OR smart 😝

Zense;
This is a very good point and not something I had considered. Thanks! It is well-known that the AI doesn't take enemy hero threat area nearly enough into consideration when sending any of its heroes off.

What you're saying is that addressing this would essentially kill two birds with one stone since this non-optimal AI behavior a huge problem in both higher difficulties and in lower difficulties.
As for combats, we haven't really made much in the way of adjustments in the AI behavior according to the difficulty level. It will eventually be done, the only reason Adventure Map behavior has been more prioritized is because you can use auto-resolve to at least get decent results. But I understand that not getting to use your troops yourself can remove a lot of the fun.
Just to double-check, you're using version 1.0.12, @Stisen11 ?
Also, if you don't mind, I'd like to share your message above in the discussion.

Stisen:
I am using the v.1.0.12 both on my PC and on my sons Android Tablet

[zenseii]

I'm adding this part of the discussion too:

Stisen:
Oh yeah! I forgot to mention, that newcomers don't consider using more than a hero or two, because it's too overwhelming to play out more than a few different calculated moves over a few turns period. My son basically forget about his second hero. He forget what he send him out to do. He doesn't see the long term benefits of using more than one hero. I remember thinking the same way, when I played hot seat with a few friends, when we were at the age of 10. It felt tedious and pointless! For some reason, we thought that walking around the map with one invincible hero on the hunt for control over 10-16 castles was the only way to play the game. Maybe we looked at the wimpy AI heroes, that walked around with almost no army and died in an instant. And when we actually did have like 4 heroes or so, we thought that the only reasonable thing to do, was to have ALL units on one hero, and NOTHING on any other. THAT'S the way a newcomer play this game, in my experience 😉 Even my grown up friends, who tries it their first time, plays like this
Zense
Yes, and I can confirm that this is the same way I played first, which is why I'm suggesting this. In my opinion, once you've learned and developed your skills to evolve past using just one hero, you're ready to play normal difficulty.

Stisen
And just like when you learn to play chess, you don't see the true potential of the pawns and see how their value is actually equal to any other chess piece, if you wanna win the game that is

[zenseii]

I will come back with some thoughts about this once I have time to sit down and think about it and organize my thoughts, not to mention look back at what @ihhub and @idshibanov have mentioned in discussions we've had about this before, which was essentially tying specific AI behavior to specific difficulty levels, like hit and run tactics to a specific difficulty level etc.

[Stisen1]

I'm not a programmer myself, so I might be naive on this matter. I have a basic understanding of limitations, when it comes to complex structures of a given code, so I would assume that most things are hard to implement, as we continue down the pathways of fheroes2 code.

Is it possible to have the easy AI player underestimate the power of a hero's army strength and hero's level? Say it NEVER takes the hero level into account, when it wants to attack. And say it can't read numbers above 'horde' stacks. It will do risky moves late game. To make it harmless early game, maybe we could make it impossible for it to attack, unless your heroes are within a certain range, and add some more criteria to this logic, so that we don't break maps, where you are supposed to get beaten by certain heroes (for instans maps where necromancer's threaten three sorceress towns in the Roland Campaign, or on the map "King of Death", where both green and blue player has a AI only hero nearby from the get go... We don't wanna destroy the intended theme and events for any given map :) For sure, this will not be an easy task...

[zenseii]

@oleg-derevenetz

There is no "hit&run" tactics for AI, at least in a conscious form.

There seems to be a misunderstanding. The hit and run tactic we're referring to is the one where you send a hero with a strong spell repeatedly to an enemy hero who is close enough to your castle that you can attack, cast the spell, retreat/surrender, and repeat, to reduce their forces significantly before the upcoming battle.

This complex strategy still needs to be implemented, so we are talking about future improvements to the AI, which is inevitably tied to the planning and design of the difficulty levels. Their idea is that such AI tactics could or should be restricted to specific difficulty levels. This was just one example of such a tactic.

[oleg-derevenetz]

The hit and run tactic we're referring to is the one where you send a hero with a strong spell repeatedly to an enemy hero who is close enough to your castle that you can attack, cast the spell, retreat/surrender, and repeat, to reduce their forces significantly before the upcoming battle.

I'm just saying that AI is not using such tactics consciously right now :)

This complex strategy still needs to be implemented, so we are talking about future improvements to the AI, which is inevitably tied to the planning and design of the difficulty levels. Their idea is that such AI tactics could or should be restricted to specific difficulty levels. This was just one example of such a tactic.

Since this tactics have not been implemented yet, there is no "lever" for it :) Future improvements for AI "maybe will be someday" (nothing like this has ever appeared in the original HoMM2/3 BTW), and it is necessary to somehow regulate what is already "here and now".

[HubKing]

the map "King of Death",

Is it even possible to win in this map? It starts at a castle that has south and east passages. On the first day, a Red hero comes near my castle from south. If I buy all the available troops and try to go to east, and my hero meets the Red hero, he gets killed. Even if I got lucky and the Red hero does not come to my hero, the east road is blocked by a feral troops, which if my hero attack, kills my hero. I can't go to south, because the Red hero comes from there. If I don't go out and stay in the castle, there's no stone, so I can't build/upgrade buildings and in about a week, Red hero comes with a few bone dragons to my castle and kills me.

[zenseii]

@HubKing, yes it can be beaten, even on impossible:
https://youtu.be/22gnBTYFCkY?si=rMN3jqu65F_9FC-N

[LeHerosInconnu]

Hello @zenseii and @Stisen1,

We can further reduce the gold income and/or limit the number of heroes on the adventure map.

Easy mode:

• -50% gold income
• max 2 heroes on the adventure map for S map
• max 3 heroes on the adventure map for M map
• max 5 heroes on the adventure map for L map
• max 7 heroes on the adventure map for X-L map

The 50% reduction in gold income may be a bit severe, but eventually.

[zenseii]

Hi, @LeHerosInconnu. I agree about setting a limit of max heroes.

There are some things I should point out though. We already have in place a system that limits the amount of heroes that the AI will recruit, and it takes into account the size of the map. And if I'm not mistaken the AI will not hire more than 4 heroes on any difficulty level for any size of map, though I'm sure we all know that it can have more heroes if the map starts it out with more. There could also be cases when a map starts the AI player off with more than the potential hero limit would allow. There is already some logic in place for this. Then there are heroes that are set to patrol or just standing still, which perhaps should not be counted.

As for cutting the gold income by half, I think we should take into account that reducing the amount of heroes in the AI's possession will effectively reduce its gold income quite a lot since it won't be picking up resources and chests, and flagging mines as quickly as it would if it had more heroes.

[LeHerosInconnu]

@zenseii,

Why limit the number of heroes?
What's the purpose, the reason?
Let AI express itself fully. :)
The number of heroes should be a function of what there is to do on the adventure map, not an arbitrary limitation.

[zenseii]

@LeHerosInconnu. As you pointed out above, the AI castles often have unrecruited creatures when you capture them, and this shows that the gold reserves aren't the best. I don't know why this limit was put in place back then, but surely stopping the AI from recruiting 8 heroes when in general they aren't needed, is a good enough safeguard against it wasting lots of gold on heroes that for 5/8 of those heroes it will send straight towards your stronger heroes to die, only to waste 2500 gold on recruiting one or several heroes the next turn.

As you said, if there is stuff to do with 8 heroes then you should recruit that amount, and right now the AI would not be using 8 heroes effectively at all, not to mention it doesn't know when to dismiss heroes, which could quite easily give it a handicap when it can't recruit heroes somewhere where there is an urgent need.

Anyways, the limitation is effective enough as a general assumption of how many heroes a human player would have on maps of those sizes. Rarely would you see a player recruit 8 heroes on a small map.

[LeHerosInconnu]

@zenseii,

@LeHerosInconnu. As you pointed out above, the AI castles often have unrecruited creatures when you capture them, and this shows that the gold reserves aren't the best. I don't know why this limit was put in place back then, but surely stopping the AI from recruiting 8 heroes when in general they aren't needed, is a good enough safeguard against it wasting lots of gold on heroes that for 5/8 of those heroes it will send straight towards your stronger heroes to die, only to waste 2500 gold on recruiting one or several heroes the next turn.

As you said, if there is stuff to do with 8 heroes then you should recruit that amount, and right now the AI would not be using 8 heroes effectively at all, not to mention it doesn't know when to dismiss heroes, which could quite easily give it a handicap when it can't recruit heroes somewhere where there is an urgent need.

Okay, I see.
AI therefore still needs to improve. :)
Eight AI heroes on an X-L map with different castles wouldn't go amiss.

Anyways, the limitation is effective enough as a general assumption of how many heroes a human player would have on maps of those sizes. Rarely would you see a player recruit 8 heroes on a small map.

Just as an anecdote, I've seen players in let's play videos recruit a plethora of heroes to have a bigger army at the start of the game.
Not necessarily a good nor a bad strategy. :)