AI Fights Well only when Already Winning

[DilithiumDad]

We all know how poorly the AI plays. This one of the major disappointments of the game. On the other hand, the AI on occasion seems to play very well and uses good strategies and builds facilities and secret projects. These flashes of brilliance from the AI are much more likely to occur early in the game. By the endgame, the AI factions have become nearly inert, and they stop building new bases or facilities and mainly churn out obsolete units but never use them for attack.

I have developed some single-player and multiplayer scenarios that include AI factions which are highly developed and tough to beat, even for human factions working in teams. On ACOL, these have included "AI's Revenge" 1 through 4, culminating in Dune1 and Dune2, both of which lead to defeat of 6 allied human factions by Deirdre and Morgan (recast as the Fremen and House of Artredes).

The secret, it seems, is for the AI to start out with the greatest might. The larger the lead the AI has in the power bar, the more intelligent its strategy and the greater its aggressiveness. One can tap into this by giving the AI extra bases and tech at the start, and pre-terraforming their territory with forest to prevent the AI's usual terraforming blunders. It also helps to have two AI factions with unbreakable pacts with each other. On ACOL, there is another series, "Hive in the Middle", in which Yang has control of all landmarks and starts out with more bases and population than all the human factions put together. Even with 7 human factions cooperating, he has not yet been beaten.

In Single-Player, I often notice superior military strategy from the AI when I am far behind in might early on in the game. Examples are the well-known Spartan Impact Rover Rush or the Believer 4-3-1 blitz. This is especially true when playing a come-from-behind faction such as Drones, Believers, or Morganites which are low in relative might in the early game and then take off in the middle game. Once you are two times stronger than the leading AI faction, the AI factions are supposed to gang up on you, according to the Prima guide. It is true that they will break their alliances with you at that point and start Vendettas. But they also stop srategizing, building colony pods, terraforming or otherwise doing anything intelligent.

My question is this. Doesn't it seem like there is some simple error in the SMAC program, where the AI is SUPPOSED to play more aggressively and strategically as the human player gets stronger, but in fact the exact opposite happened? Maybe two lines of code got switched somewhere. It seems that the AI really comes into its own when its might is two times that of the human player. That is the exact opposite of what is supposed to happen!

So, is this a simple bug in the programming code? Any word from Fireaxis on this? Will this absurd behavior be fixed in Civ3?

[Misotu]

Yes, this is a good question. I don't have the answer, but it's a good question

Having played AI's Revenge, Dune and Hive in the Middle, I can certainly confirm everything that you have said. Although the humans will certainly beat the Hive easily in Hive in the Middle 3. But having said that, the interesting aspect of the HIM games is what the 6 human players will do when they've destroyed the Hive

I think you'd probably agree though that there are limits to manipulating this "feature" in order to obtain a good game from a scenario. It's true that the stronger it is, the better the AI plays. But I've found that tweaking scenarios to make them harder and harder is not generally so very satisfactory in the end. They tend to be far too hard in the early days - with synth scouts facing chaos jets for example - but then tail off again as the human player develops (assuming the human survives the early mayhem).

What we really want is an AI that can do what a human opponent does: start off on an even footing, and stand a good chance of winning!

Don't know what the answer is. For me personally, the answer is to play multiplayer games, rather than solo games against the AI which no longer interest me very much. But it would be good if Civ3 came up with something better - I used to enjoy my solo games. The great thing is that you don't have to wait for the next turn to arrive

[Lefty Scaevola]

I have seen the AI (planned? accident?) use artillery in connection with a probe swarm. Deidre bombard former base which I had captured with artillery first, severely weaken the unarmored and thus vulnerable probe teams I had there, then killed them with her probe and then mind controlled the base. Previously she had use a swarm of infantry to attack supported by artillery that I had just managed to beat back. This was at transcend level I was number 1, she was number 3.

[DilithiumDad]

This is the sort of flash of brilliance I was talking about. One thing the AI does consistenty well (better than human players, actually) is bribing back captured bases. This is a strategy I have copied fomr the AI to good effect --your own captured bases can be bribed very cheaply.

The question is, why doesn't the AI show this sort of strategizing more often?

[SerapisIV]

It's been pretty much assumed that the AI's intelligence is based on their power ranking. Do we know for sure that this is true? Does anyone know where in the code that the power formula exists? Maybe it can be hacked, maybe inverse all the power ranks. That way the AIs might get more intelligent, especially if they think the player is super-weak.

A possible problem would be the factions not submitting after they get whipped on, but by the time they get that weak, it really doesn't matter anyway.

[Flubber]

I have not really witnessed better strategy from the AI when it is more powerful. It just seemed to me that when the AI is the biggest and the baddest, their basic predictable straight ahead strategies work well through the sheer force of numbers and tech superiority.


It seems to me that the AI will still dribble units toward you rather than concentrate for a concerted attack but if the AI has three times as many bases as you and tech superiority that dribble becomes serious. If as you suggest there is some function by which the AI gets better in certain circumstances I wish they would just turn it off and have the AI play its best always.

I really just think its the difference between being at odds with a 100 lb stupid guy and a 6 foot 8 280lb stupid guy.

[Scipio Centaurus]

Originally posted by DilithiumDad
This is the sort of flash of brilliance I was talking about. One thing the AI does consistenty well (better than human players, actually) is bribing back captured bases.

If you are talking about the AI using probes teams to mind control recently captured bases away from their human captors, I agree. Seems like in the late game, every time I clear out an enemy's garrisons and take a base with drop troops, the AI automatically tries to mind control the base back with a probe team on the very next turn.

I've gotten into the habit of actually moving the invading drop troops one square out of the base in the direction(s) I know the probe team(s) will be coming from (just to block the probe teams) and then immediately rush building a probe team of my own in the newly captured base. Although I lose a fair number of drop troops that way, I keep a much higher percentage of captured bases.

- Scipio

[Googlie]

I've been experimenting recently with a number of scenarios with master/slave relationships.

e.g. Democracy in Peril and The Axis Awakens. Premise is that a strong faction, having subjugated the research factions (University and Conciousness) will be hard to beat. they are set up as human collaboratibe games, or as SP games. I got my ass whupped in both as SP players.

Am currently working on scenarios (which will soon be posted in Rynn's Gallery) where there are two competing camps, each headed by an alien faction with 2 slaves - a builder and a researcher. the seventh faction (you) must choose which one to support. Am having great fun right now as the Drones against a H'minee/Deirdre/Aki bloc and a Marr/Zak/Morgan bloc.

But I agree - it shouldn't be necessary to tweak through scenarios (although if enough of us built a dozen each, and posted the starts somewhere, there'd be a huge library of challenging SP games)

Rynn's Gallery has a few - it can be found at:

http://googlie2.tripod.com/rynnsgallery/

The Roze and Morgan CGN Challenge ones were fun to build and play.

Googlie

[Basil]

The Alpha Centauri manual talks of how they deliberately designed the AI to "[amass] an overwhelming force and then move up to attack one of your bases". Though this is a good tactic in the rover rush years, it's not so good after air power is developed: keeping your high attack-one defence-high movement air units back until you've amassed an overwhelming force, while ignoring all targets of opportunity and local superiorities , is very poor strategy, particularly when the player has choppers that can kill a lot of one-defence aircraft each turn.

I've found that disallowing crawlers and pop booming helps keep the computer in the game longer and makes the game more interesting in general.

[Avenoct]

Where's that Might Thread?

Good discussion so far. Does anyone remember the AI might thread from a couple months ago? Maybe we should reference it here. In that thread there was discussion about granting all AI factions a tech with a huge 'AI-Influencer' that otherwise did nothing. The AI factions would then believe themselves mightier, and hence, play more intelligently.

I think though, that we can trace the issue back to the flawed strategy represented by:

In Single-Player, I often notice superior military strategy from the AI when I am far behind in might early on in the game -D-Dad

and

The Alpha Centauri manual talks of how they deliberately designed the AI to "[amass] an overwhelming force and then move up to attack one of your bases". - Basil

Added together, we get a basic AI strategy of: "attack when the ratio of 'My Forces / Their Forces > 1."

This is flawed because, against a superior opponent, waiting untill you can utterly defeat it is not a probable occurence. Think about how a human gets out of this situation. I might say "Against a superior AI I will wait until I have sufficient forces to secure a single AI base, and then attack that base." Overly simplistic, but very context specific. I believe the AI could be written to take context-specific cues like this, but this is hard-wired to the game-engine, so is unchangeable unless Firaxis releases the uncompiled source.

But what happens when (Myforces / Theirforces > 1), as in, when the AI (Myforces) believes it has the upper hand? The AI then attacks, and does so without much concern, which turns out to be fairly decent behaviour.

Adding Might and Diplomacy into the formula we might get: Attack if ((Myforces * Mymight) + MyDiplomaticstanding) > ((Thierforces * Theirmight) + TheirDiplomaticstanding)). Where Might is obviously the calculation referred to in the help-text and Mightgraph, and Diplomaticstanding might be the ratio of Allies to Enemies.

So no, I don't think the code was mixed up, it was just never made context-specific and uses a formula that only makes sense when the AI is ahead.

-Smack