How Should Beyond Earth Handle AI Difficulty Scaling?

[Matthew.]

If you don't give the AI initial bonuses then the human will just cripple the AI at the start.


I think it's really hard to say what bonuses the AI in CBE needs since I know so little of the actual game mechanics. In Civ5 however I would have preferred if the AI got straight up combat bonuses (a strength 10 unit would be strength 12) at the higher levels. It's a lot more in-your-face than the economic bonuses the AI does get, but also a lot more directed at where the AI actually needs help. The AI in Civ5 doesn't need more units, it needs better units.

Oops, didn't even see this post. I agree

I play a modded version of 5 which does just this, and it works great. The human player needs to lose units in war, like the AI does, that is the base of the problem. Sniping wave beyond wave of enemy units puts them behind thousands of hammers, gold, time, and you lose nothing. It is indirectly pushing you further ahead just by playing relatively passively (letting AI suicide units upon walls).

I also think it helps AI war-mongers take cities (much easier to take a 50 strength city with 75 strength equivalent units than 50 strength units), so I see less of the issue of "typical AI suicides all units 50 turns into the game, now sits on the sidelines 3 eras behind everyone else". However, we don't know how the city mechanics will work in BE, so this could very well be completely irrelevant. But I still stand firm on the point that, in combat, the human player needs to lose some military units like the AI does, or it snowballs into a huge advantage by mid-game.

 

[AriochIV]

http://jonshaferondesign.com/2012/09/10/the-recipe-for-good-ai/ - that's a post by Jon Shafer, original lead of Civ 5. He also have some additional info further in his blog.

Shafer's article on the pitfalls of AI is quite good; I guess he should know, as the initial release of Civ V failed on every point that he mentions. (Note that he hints that it was a lack of proper coordination between himself and the AI programmer (Ed Beach) that was responsible for these failures.)

However, also note that nowhere in the article does he mention lack of CPU resources as a source of the problem.

I'd also like to note that somebody at Firaxis must know a thing or two about AI, because the tactical AI in XCOM is pretty good. It's one of the few examples in recent memory in which the AI actually plays smarter at higher levels, instead of just stacking on bonuses.

 

[stealth_nsk]

IMO enemy combat units should be given overall buffs, increasing on difficulty.

One of the worst suggestion ever. This form of cheating is most irritating.

Shafer's article on the pitfalls of AI is quite good; I guess he should know, as the initial release of Civ V failed on every point that he mentions. (Note that he hints that it was a lack of proper coordination between himself and the AI programmer (Ed Beach) that was responsible for these failures.)

The Civ5 released was rushed. It lacked so many things, that it's difficult to blame developers for the AI.

However, also note that nowhere in the article does he mention lack of CPU resources as a source of the problem.

The blog is about game design, not programming. You can't cover everything.

I'd also like to note that somebody at Firaxis must know a thing or two about AI, because the tactical AI in XCOM is pretty good. It's one of the few examples in recent memory in which the AI actually plays smarter at higher levels, instead of just stacking on bonuses.

XCOM is absolutely different thing. You don't need to calculate each tile there - the only things which matter are covers, line of sight and distance, all with very limited number of soldiers. If you define conditions right, you could implement brute-force here, the thing which is not possible in civ.

 

[Trias]

One of the worst suggestion ever. This form of cheating is most irritating.

Why? Why is this worse than the AI not paying unit maintance and building units at twice the rate?

Tha AI "handicap" settings should effectively ameliorate the weaknesses of the AI. This is one of the points where the AI design of civ5 failed big time. On one hand they designed the game to favor quality over quantity. On the other hand, they tried to compensate the weaknesses of the AI by giving it more units and more cities. This obviously is very ineffective. Consequently, the AI bonusses have to be screwed way up to immersion breaking levels.

It would have been much more effective to have all AI units start with a free promotion (e.g. "Emporer level" "Diety level") that gave small boosts to the combat strength than have the AI build twice as many units. (Add advantage would be having less units to move and therefore shorter turn times) The AI probably would still need a little help from production and upkeep bonusses, but not at the present level.

At the highest level the bonuses of on the promotion should be ramped up and could also include:
-bonuses to movement (helps the AI resolve log jams)
-bonuses to sight range (helps AI make better informed decisions)

 

[Kid R]

Well, that's quite a small part of the post ...

Yes, definitely! That was just the bit that got me thinking right then. Soren Johnson's talk should be required viewing for anyone discussing this stuff

 

[Kid R]

I'd also like to note that somebody at Firaxis must know a thing or two about AI, because the tactical AI in XCOM is pretty good. It's one of the few examples in recent memory in which the AI actually plays smarter at higher levels, instead of just stacking on bonuses.

I never noticed the XCOM AI get smarter through the game. The different enemies behave differently so maybe they seem smarter, but then the sectoids being dumb and the etherials smart makes total sense in the context. It even makes sense for sectoids under mind control by etherials to be smart. It wouldn't make sense for Gandhi in Civ to start dumb and become smart later though. Although sure it doesn't make sense for him to live for 4000 years either

Edit: it could make sense for great generals attached to AI units to enable extra smart tactics for units round them. Like a 4 star general would have better tactics than a 2 star. So the human player could know this and see them coming and go "Uh-oh, here comes General Rommel, he knows the pincer technique" or whatever...?

 

[Mustakrakish]

On one hand it does kinda sounds like a good idea to just buff their military quality wise, being either higher overall strength or a bunch of free promotions and a bigger combat bonus to Great Generals (additional +10% or something), and while I see it clearing the map and confusion a bit (no carpet of doom etc.) and making for more interesting wars (vacuum cleaning the carpet isn't interesting)...

I see on the other hand this creating frustration and inferiority complex in players in the long run, trying to much army of all those around you, as players army will always (or almost always) be inferior. While outwitting and using superior tactics will be fun to overcome superior units instead of numbers, someone is bound to get pissed at the idea that his units are always inferior.

I really can't see any win-win solutions to the whole deal, except for removing bonuses altogether and improving AI itself, but I'm not qualified to add my thoughts on that.

 

[Trias]

I see on the other hand this creating frustration and inferiority complex in players in the long run, trying to much army of all those around you, as players army will always (or almost always) be inferior. While outwitting and using superior tactics will be fun to overcome superior units instead of numbers, someone is bound to get pissed at the idea that his units are always inferior.

There are two factors counteracting that:
1) Human player in general will be better at keeping units alive. Consequently, the human's units will be better promoted. (This is one AI weakness that is ameliorated by given AI units a (slight) buff.
2) The Human will generally have a tech advantage giving him access to better units.

These two factors all the human to somewhat "level" the playing field against the AI.

As with any AI bonus, the bonus should never be too big. (e.g. have AI opponents with warriors that cannot be killed by your spearmen is very obnoxious) Boosting both AI unit quantity and quality a little is probably more acceptable to most players than boosting one of them by a lot.

I really can't see any win-win solutions to the whole deal, except for removing bonuses altogether and improving AI itself, but I'm not qualified to add my thoughts on that.

That would not solve the difficulty level problem. Writing two versions of the AI, where one is consistently "dumber" than the other is possibly even harder than improving the AI. (let alone 5 different ones) IMHO that would be a total waste of developer resources. I'd much rather they invest that in making the AI really good, and differentiate between difficulty levels by AI bonuses and handicaps. Ideal would be if diety is the "level playing ground" level, and all other levels giving the human a leg up, because the AI is that good. But that is probably never going to happen.

 

[JokerJace]

That would not solve the difficulty level problem. Writing two versions of the AI, where one is consistently "dumber" than the other is possibly even harder than improving the AI. (let alone 5 different ones) IMHO that would be a total waste of developer resources. I'd much rather they invest that in making the AI really good, and differentiate between difficulty levels by AI bonuses and handicaps. Ideal would be if diety is the "level playing ground" level, and all other levels giving the human a leg up, because the AI is that good. But that is probably never going to happen.

You would not need to write two versions of the AI. You could simply implement clever parameters that are difficulty based and determine the likeliness of certain decisions. For example how many information are taken into account. Or simply vary the probability distribution of a nondeterministic algorithm. Modular Software Engineering!

 

[AriochIV]

There are certainly fewer units in an XCOM battle than on a Civ V map, but otherwise, the algorithmic problems are similar; they're still units on a grid that can't share the same space, have a limited number of movement points each turn, and must maneuver in relation to hostile units. If anything, the mechanics that must be considered are much more complex in XCOM than in Civ V; each unit has different weapons to choose from, limited ammunition, variable chances to hit based on distance and terrain, morale, etc.

I never noticed the XCOM AI get smarter through the game. The different enemies behave differently so maybe they seem smarter, but then the sectoids being dumb and the etherials smart makes total sense in the context. It even makes sense for sectoids under mind control by etherials to be smart. It wouldn't make sense for Gandhi in Civ to start dumb and become smart later though.

I meant the AI gets smarter on harder difficulty levels. On Normal, the AI is more passive, alien units often staying more or less where you discovered them, and sometimes doing nothing when it could take a key shot at an exposed unit. On Classic, aliens move much more aggressively to flank you, are more likely to pull back and go on overwatch when in an unfavorable situation, and rarely miss a chance to make you pay for a reckless move.

 

[Trias]

You would not need to write two versions of the AI. You could simply implement clever parameters that are difficulty based and determine the likeliness of certain decisions. For example how many information are taken into account. Or simply vary the probability distribution of a nondeterministic algorithm. Modular Software Engineering!

Yes, sounds nice in theory. Now apply that to a decision making routine for moving units such that you get 5 different AIs in equal increasing steps of difficulty.

 

[AriochIV]

There's another aspect to this issue that's specific to Beyond Earth: the AI factions arrive on the planet much later than you, so the AI is going to have to receive bonuses to catch up with you, even on the easiest difficulty levels. The Beyond Earth AI is going to have to cheat, by definition.

 

[stealth_nsk]

There are certainly fewer units in an XCOM battle than on a Civ V map, but otherwise, the algorithmic problems are similar; they're still units on a grid that can't share the same space, have a limited number of movement points each turn, and must maneuver in relation to hostile units. If anything, the mechanics that must be considered are much more complex in XCOM than in Civ V; each unit has different weapons to choose from, limited ammunition, variable chances to hit based on distance and terrain, morale, etc.

The number of choices is different. You could remove all invalid choices right away. In the first proximity:
- Melee aliens just walk to nearest enemy.
- For ranged aliens tiles in cover from all enemy units are the only valid choice. Tiles which allows shooting are preferred, tiles which allow flanking are better.
This logic requiring nearly zero implementation efforts provides already valid AI for movements.

For civilization this will just wouldn't work.

 

[Mustakrakish]

There are two factors counteracting that:
1) Human player in general will be better at keeping units alive. Consequently, the human's units will be better promoted. (This is one AI weakness that is ameliorated by given AI units a (slight) buff.
2) The Human will generally have a tech advantage giving him access to better units.

These two factors all the human to somewhat "level" the playing field against the AI.

As with any AI bonus, the bonus should never be too big. (e.g. have AI opponents with warriors that cannot be killed by your spearmen is very obnoxious) Boosting both AI unit quantity and quality a little is probably more acceptable to most players than boosting one of them by a lot.

Yes, now that I think of it, there's no point frustrating about it since one chose level of difficulty.

That would not solve the difficulty level problem. Writing two versions of the AI, where one is consistently "dumber" than the other is possibly even harder than improving the AI. (let alone 5 different ones) IMHO that would be a total waste of developer resources. I'd much rather they invest that in making the AI really good, and differentiate between difficulty levels by AI bonuses and handicaps. Ideal would be if diety is the "level playing ground" level, and all other levels giving the human a leg up, because the AI is that good. But that is probably never going to happen.

Yes I should've clarified that in that case (with competent AI) I don't see reasons to make several levels of difficulty, except maybe giving bonuses to players at easy levels to get into the game. Or some simple bonuses to AI if one wants a serious challenge and even competent AI is not enough.

 

[Trias]

The number of choices is different. You could remove all invalid choices right away. In the first proximity:
- Melee aliens just walk to nearest enemy.
- For ranged aliens tiles in cover from all enemy units are the only valid choice. Tiles which allows shooting are preferred, tiles which allow flanking are better.
This logic requiring nearly zero implementation efforts provides already valid AI for movements.

For civilization this will just wouldn't work.

Also, X-com doesn't need a generic AI that will handle any unit that can be specified through xml data. This means that for X-com programming (semi)-individual strategies for all alien types is an option.

 

[Trias]

Giving the AI omniscience or built in maphack is a good example of an advantage that many players find unfun and overly cheap.

I think giving the AI forms of omniscience is unnecessarily frowned upon. Whether omniscience is annoying to the player depends on how the AI uses it. Let me illustrate by example,

If the AI can see the remain build times of other players on wonders and uses this to rush the completion to beat the human player to a wonder, than that is incredibly annoying.

However, if the AI at higher level uses the remain build time of other players when deciding to start building the wonder in the first place to gauge whether he has a chance of completing it first, then that is not quite so bad. It actually simulates game knowledge on the side of the AI quite nicely. High level human players typically know around which turn wonders get completed and factor that into their decision to build them or not. This information is typically not available to the AI (for one the typical turn of completion will change as human strategies evolve, not to mention that it will depend on map type etc.). Using this omniscient knowledge in this way does not hurt the player, it merely prevents the AI from making mistakes that hurt it in the long run.

A middle ground would be to give the AI access to the remaining build times plus or minus a random number of turns to simulate that it is making an "educated" guess.

This applies more generally to the AI using information that strictly speaking it should not have access to. If done right, it can provide an efficient way of simulating AI knowledge of the game (that would otherwise would be impossible//impractible/computationally expensive to code).

 

[Rakshasa72]

IMO the X-Com AI does get cheats. It gets Light Plasma Rifles way before the player can (which is a big deal). Later in the game it gets units that count in heavy cover standing in the open. More units/health based on difficulty. While the player gets less health based on difficulty. I think I read somewhere there is an underlying mechanic that will force hits/misses based on difficulty settings helping to break streaks. So yes even X-Com AI gets cheats.

 

[Lord Tirian]

So yes even X-Com AI gets cheats.

Of course, there's a big difference: Civilization is about AI civs starting with the same stuff you have. Apart from their unique abilities/units/etc., they're equals. The AI is a competitor first and an adversary second.

In XCOM, the AI is playing a hostile invading force at a higher tech level, the deck is stacked against you on purpose. The AI is an adversary and not a competitor at all.

In other words, in XCOM, the AI is the mountain you climb. In Civ, the AI are the other marathon runners you try to beat.

 

[Rakshasa72]

@Lord Tirian

I guess my point was more that even in X-Com the AI gets static bonuses such as increase health and accuracy at increased difficulties. So I guess that even if the AI does play smarter at higher difficulties (I think it does). They still feel they need to give it hacks similar to the hacks they give Civ AI.

 

[Kid R]

I meant the AI gets smarter on harder difficulty levels...

Ah OK, I never saw that - only played it on standard difficulty!

...
In the first proximity:
- Melee aliens just walk to nearest enemy.
- For ranged aliens tiles in cover from all enemy units are the only valid choice. Tiles which allows shooting are preferred, tiles which allow flanking are better.
...

In general how's this: 2 separate difficulty levels. Economy difficulty would be the gold and hammer bonuses we're used to from other games. Combat difficulty would gradually add in a sort of library of tactics. E.g. "Private" = units just walk towards nearest enemy. "Corporal" = units use terrain. "Sergeant" = AI builds army composition tailored to its war target. "Captain" = Uses suicide worker decoys or whatever

Watching the AI and and working up through those levels could teach new players the basics of how to handle the units in the game, as well as making it harder in an honest way.

Also a benefit of escalating the difficulty via 2 separate sliders is that it would let warmonger and peacemonger players tailor things better to how they like their games to play out.