Some comments on the AI Difficulty levels

[Fulano5321]

I made this post on Facebook a long time ago and the moderator suggested posting it here.
Weeks later I finally got the forums to actually let me register so here goes.

Could you have more levels of AI difficulty where the AI doesn't heavily cheat?

It's a problem in a lot of the 4x games I play, the AI is dumbed down through the normal difficulty levels, but at higher difficulty it gets smarter and gets huge bonuses at the same time. The smarter AI makes the game a lot more interesting to play, but the added cheating at the same time makes a very large difficulty jump.

For example, the AI in Age of Wonders 3 the AI plays very tame at Lord difficulty. At Emperor I see much smarter tactics such as kiting and casting the highest level spells, but their huge bonuses leave the player doing a lot of tedious gorilla type tactics, because each AI is basically two players worth of bonuses.

I agree that it's good for an AI to get some income bonus so they can throw stacks of units away at the human to keep them entertained, but in this game the player has to make major changes in tactics to be able to deal with the AI's huge bonuses.

Could you add AI difficulty levels in Planetfall where it plays it's best with only mild cheating, then continue adding more cheats at higher difficulty levels to make the game harder for those who like overwhelming odds. I'd love to see such a feature added to Age of Wonders 3 as well.

 

[BloodyBattleBrain]

You mean like 2 sepa rate option sliders.

Slider 1:

Squire ---> emperor

Slider 2

100% bonus ----> 100% malus


Personally I don't mind cheating ai as much as I used to, as long as it is fun.


I believe the AI in AoW3 follows 1 or 2 scripts, which is basically turtle then overwhelm with numbers and high tier units.

I could be very wrong though as I'm not an expert or a programmer or a pro gamer.

But my experience indicates Warlords and dreads are dangerous AI emperors but Necromancers are really not.

I suggested that each class have its own script, it's own instructions for playing the game.

And that there be sub routines.

For example, the Napoleon AI would be based on dread with massed musketeers and cannons (perfectly valid strategy, and with emperor bonuses even more so. Would be Goddammed scary getting rushed by 4 to 6 stacks of musketeers by turn 25 or 30 and then follow up stacks with Cannons.)

Alexander would be masses of pikes and then phalanxes with Cavalry.

So Human Warlord would be the Base.

Obviously all could be applied to any race.

Attila would be lots and lots of horse archer + support spells abuse (lots of seeker, haste)

Stalin would be a Necromancer with many many cadavers and then massed bone horrors to push through.

I am led to understand it'd be a bitch to program, to put it mildly.

Obviously they'd have to be back up routines otherwise if you know you're facing Stalin you just stock up on support units.

In which case an algorithm like: if x% of enemy is support units, switch to massed cavalry.


There'd be variations of each flavour. So being an Elf would modify Napoleon to use seeker spells.


Also, maybe the AI would be randomised.

So you choose emperor and it goes through and picks an AI profile.

That'd keep things a bit more interesting.

 

[HousePet]

I tend to feel that AI cheating/bonuses tend to overshadow the AI actually being better with most games.
That causes turning the difficulty up to mean that you just end up targeting flaws in the AI, which makes the difficulty feel hollow.

 

[The Founder]

Could you have more levels of AI difficulty where the AI doesn't heavily cheat?

I tend to feel that AI cheating/bonuses tend to overshadow the AI actually being better with most games.
That causes turning the difficulty up to mean that you just end up targeting flaws in the AI, which makes the difficulty feel hollow.

Game AI has the planning ability of a falling Slice of Pizza. All AI does. They run always in the "here and now", with no concept of a past or a future.

Any appeareanche of planning ability is just a well crafted illusion. It would be a huge step forward for AGI development if we could model memory or planning. We kind of hope that deep learning networks will evolve that on it's own.

So we will not get around AI cheats in any strategy game. Ever.

 

[HousePet]

That is very wrong. There has been planning AI in these games for a while. They just tend not to be good at playing the game.

 

[The Founder]

That is very wrong. There has been planning AI in these games for a while. They just tend not to be good at playing the game.

Any examples that actually apply to Economics and Military AI, the two most needed ones?

 

[Fulano5321]

You mean like 2 sepa rate option sliders.

Slider 1:

Squire ---> emperor

Slider 2

100% bonus ----> 100% malus


Personally I don't mind cheating ai as much as I used to, as long as it is fun.

That could work, so a player could work up to the hardest AI, then start working up the bonus they get. People might like that because they could directly see what bonus the AI is getting.
Personally I was thinking just adding a couple more notches and reduce the cheating on King and Emperor, but I do like the idea of having two sliders better.

I seem to remember doing that in other games... except it was giving myself some sort of handicap after the AI was maxed out.

Personally I don't find it fun to be overwhelmed with so many units. It's pretty easy to pick off the AI's units as it's moving them from point A to point B, but I don't see that as a reason to just give them 2x the units. On Emperor I also never get the research awards, I don't really feel like there is much of a chance to unless you happen to get a really good start. Usually what happens is I don't kill the ai's fast enough and end up with 2-3 of them teamed up against me. At that point I might as well just say goodbye to having any sort of word spells running and I have to just slowly gorilla tactic my way through their empires one at a time.

 

[Leon Feargus]

There have been many complaints and lengthy discussions about the AI in aow3. I think the general opinion (may be wrong here) is: The AI is bad, but not as bad as in many other games. One thing I have come to understand is that programming the AI is the most difficult part of the whole development process.

You mean like 2 sepa rate option sliders.

Slider 1:

Squire ---> emperor

Slider 2

100% bonus ----> 100% malus

This is a great idea actually and possibly not too difficult to implement.

Game AI has the planning ability of a falling Slice of Pizza. All AI does. They run always in the "here and now", with no concept of a past or a future.

Any appeareanche of planning ability is just a well crafted illusion.

You may be right here but if the illusion is crafted well enough then that´s good enough for me.
Could you explain how a chess AI works? Because it sure seems to be planning ahead.

 

[The Founder]

You may be right here but if the illusion is crafted well enough then that´s good enough for me.
Could you explain how a chess AI works? Because it sure seems to be planning ahead.

With Chess we have literal centuries of documented human play. The game is only 2 player with on average something like 20 moves/turn avalible. And then we have been trying to handcraft the algorythms longer then we had computers capable of executing them. You can somewhat reliably predict if a specific boardstate get's you closer or farther away from winning in the end.

Most of the tricks used with Chess did not even scale up to Go, wich just has an order of magnitude higher branching factor. And figuring out of a boardstate get's you closer or farther away from victory is harder.
He can explain this ebtter then me:

 

[HousePet]

Civ4 definitely had the AI decide to attack you, build up an army over multiple turns and then declare war (no matter how silly it was). I obviously haven't seen the code that does that, but that game does let you see everyone's current military power, and you can load a saved game and stop all your current wars to be ready to crush them when they try to attack you, and they still try it. Economics, harder to tell as you can't see everything the AI is doing. Sometimes it looks like they are trying for a Space Victory, but its hard to tell if they are focussed on that or not.
Chess AI also does pretty well for a being a falling slice of pizza.

 

[The Founder]

Civ4 definitely had the AI decide to attack you, build up an army over multiple turns and then declare war (no matter how silly it was).

Or it just build an army because it was below the "recommended army for empire size, time and enemies" thresholds.
And then attacked you because it was beyond the "hate you so much, I do not care for the odds anymore"* threshold and the "truce lockout" had passed.
No planning ability needed at any spot.

*Really abysmal relations can be used for the AI to tell itself something across turns. At a certain point the AI just goes into Berserker mode. I it is propably asumed that gamestate of Relations happens if hte player is really far ahead or has caused really severe losses for the country. But in soem cases it might jsut hit that state by accident.

 

[HousePet]

Its possible that it worked that way, but then its lost the illusion of planning and is deliberately making an incorrect tactical move? Something was definitely wrong with its coding, but you are suggesting that it was deliberately coded to do stupid. Possible, but unlikely.
You can tell that it isn't just a threshold for army size, as if that was the case all empires would have power ratings equal to their size, and that wasn't the case. You could see a clear ramp up of military production, that wasn't correlated with anything else (except the declaration of war many turns later).
There was something similar to a "hate you so much, I do not care for the odds anymore" threshold, but it actually decided whoever had the lowest relations was the Hated Enemy, regardless of how nice they were. Was very weird at times.
One thing I find a little odd about these discussions is that they always seem to assume that reacting to the current situation and planning for future turns are exclusive. Since it is rarely possible to win on most turns of the game, how can it not be planning for future turns? People also seem to assume it is really hard to have the AI pick a goal and alter its per turn choices to target that.