Dominions 2. AI. Suggestions, that how to fix it.

[Pocus]

a poster got a very valid point about the AI : the AI dont go enough for quality. This is why we see so many indeps units in the AI armies.

I think the AI algorithm should be refined on what to buy. An ideal proportion of weak troops should be fixed, and recruitment in these weak troops should be stopped when the % is reached.

You can then refine the system by changing the proportion depending on what type of AI level, the type of map, or nation the AI play, but as of now, here is a raw first implementation :

troop quality (TQ) : gold + resource cost

National average of TQ : for each unit (not leader!) type in the national roster, compute TQ. Average it. It gives the average TQ of the nation (Ulm would have a higher TQ than say Ctis IMO). Lets call it NATQ

ideal TQ proportion to be reached:
TQ inf 0.5 NATQ : 20%
0.5 NATQ inf TQ inf 0.9 NATQ : 20 %
0.9 NATQ inf TQ inf 1.3 NATQ : 40%
1.3 NATQ inf TQ : 20%

have to use inf for the inferior sign, as otherwise the post decipher it as an HTML flag

Troop recruitment:
order of province browsing: to be determined smartly. I can post an algo, but this is not my point here, and it have to include a bunch of params.

so ...

get a sorted list of province where the recruitment will take place.
sort each type of troop available, from highest TQ to lowest. if troop Category is in a TQ which is undermanned, recruit 1. Repeat until no more gold, resource, or TQ is ok. Proceed to the next troop type.
Then proceed to the next province.

improvment :
overrides (partial or complete) should be used if the province has a particularly high interest in the 'province which will undertake recruitment list'.

this is very crude. You can refine by splitting desired proportion between main line units, missiles, cavalry, etc.

[Pocus]

forgot something of importance:

Some of you wonder why the tac AI is so good, whereas the strat AI is average.

The main difference lie in a word 'cooperation'. Cooperating mechanisms in AI coding is one of the hardest thing to do well. It is even more difficult that learning behaviors.

A given leader/unit in the tac AI dont cooperate that much with the other. Sure, it evaluate that casting poison clouds everywhere if your army is not poison warded is detrimental, but a simple fitness function can tell him that. The tac Ai, for spell casting for example, browse all the spells at his disposal, and get a numerical appreciation on how the spell would be of interest.
With that in mind, with some serious tweaking and work, you can get a good AI for your mages.

The situation is completely different and much more complex for the strat AI. You cant simply tell him to start with biggest stack, seek the best objective for this army, then proceed to the next. I wont discuss further the issue, but believe me the devs made already a very good job with the dom AI.

so much for the 'why dont they do a strat AI as good as the tac AI'

[Aristoteles]

Yeah the AI should be more aggressive if you are at war with him, also the AI is mainly using cheap - weak units.

Just a question: Can't you guys contact with Adam West from Crosscut about scripting? Maybe he could help you scripting a better strategical AI, if hes got enough time for that.
Just a suggestion..

[-Storm-]

Originally posted by Pocus:
forgot something of importance:

Some of you wonder why the tac AI is so good, whereas the strat AI is average.

The main difference lie in a word 'cooperation'. Cooperating mechanisms in AI coding is one of the hardest thing to do well. It is even more difficult that learning behaviors.

A given leader/unit in the tac AI dont cooperate that much with the other. Sure, it evaluate that casting poison clouds everywhere if your army is not poison warded is detrimental, but a simple fitness function can tell him that. The tac Ai, for spell casting for example, browse all the spells at his disposal, and get a numerical appreciation on how the spell would be of interest.
With that in mind, with some serious tweaking and work, you can get a good AI for your mages.

The situation is completely different and much more complex for the strat AI. You cant simply tell him to start with biggest stack, seek the best objective for this army, then proceed to the next. I wont discuss further the issue, but believe me the devs made already a very good job with the dom AI.

so much for the 'why dont they do a strat AI as good as the tac AI'

That is true.
Anyways, the devs will figure out a way to improve the strategic AI, I am sure about that.

[Zerger]

Originally posted by MStavros:
Okay I decided to make this topic, since the AI is very weak.

As I see, the biggest problems are, that the AI:

1.) isnt recruiting enough units.
2.) is trying to operate with small 'blitz' armies, I never ever seen a bigger AI army in the demo so far.
3.) isnt protecting its provinces very well
4.} isnt making smart strategical moves


Tell me what you think. I think these are the major problems with the Doms 2. AI.

1 - I agree. However some AIs had nice amount of units in some of my games, but all of those were weak units. They were toast against my heavyweight army.

2. - Very true

3. - Hm I don't agree with this. This part is working well. When I leave a province with a small defense force only, the AI always trying to conquer it with a small army, and he is getting it almost every time.

4. - I agree. One of the biggest problem is, that the AI has no counter for any sort of
efficient strategy.


2 main problems with the AI:
* The AI won't recruit an effective army. Mostly it is trying to attack with tiny - mediocre armies.

** The AI has no counter for any sort of
efficient strategy.


Setting indy str to 7++. It helps a bit, that will slow down the human player, but it also slow down the AI. Also I noticed that when I set it to 7, 8 or 9, and I reach the AI, its hellish easy to conquer the AI's provinces, since the AI had killed the indy armies, and propably the forces of the AI are almost totally dead when I reach the freshly conquered provinces.

===>> This won't help at all in anything.

[ October 29, 2003, 12:54: Message edited by: Zerger ]

[apoger]

The main issue is that humans use very focused strategies, compared to the computer that tends to dabble in a bit of everything.

Sadly also (just like Dom I) it looks like the computer doesn't use it's pretender for anything beyond sitting in the capital doing research, casting rituals, and making magic items. It doesn't use the predtender for combat advantage the way a human would. As far as I can tell, it also doesn't use the pretender for site searching (in person).

On the impossible level Illwinter should have hardcoded more focused behavior. Only allowing certain units, certain research, and certain spells to cast. It wouldn't be as good as a human, but could mimic some human behavior.


Hey IW- If you guys want I could come up with a few simple rules for each nation. Perhaps you could code that into a new level of AI.

[ October 29, 2003, 13:06: Message edited by: apoger ]

[MStavros]

Yeah, if the indy setting is 7 or more, nothing will happen. The only difference is, that you cannot expand that fast.

I think we mentioned the biggest weaknesses of the strategic AI. It is up for the devs now, to upgrade it.

Oh and Kris, if you had some games against the AI, please post your opinion.