How to improve the AI's?

[Urban Ranger]

Or you can just make the AI capable of learning from mistakes in the first place.

[Sikander]

Originally posted by Darsnan


The AI's would build them, but the units would still have movement points - no one has discovered yet how to disable movement points for units so that they are static.



There are many significant variables involved in what you are describing, a few of which I'll list below:

1. MY in game: the MY is a variable and thus effects what viable tactics the AI should employ: a rush of infantry/ NL/ speeders will work exceptionally well before aircraft, OK after needlejets are employed, but would be massacred after choppers have entered the scene.

2. Each game is extremely individualistc in that it takes on the personality(s) of the individuals involved. Different players play the game differently, and so because of this one player may arrive at a point on a map at MY22XX, however how they arrived at that location can and probably will differ significantly from another player exposed to a similar environment. This then would make any "modelling" pretty much moot.

3. The physical makeup of Planet: 30% Ocean/ 50% Ocean/ or 70% Ocean? Combined with the shapes of the available landmasses. Combined with little rainfall/ moderate rainfall/ or heavy rainfall? Combined with little NL/ Average NL/ Heavy NL, combined with the mix of which human controlled Factions versus AI controlled Factions, etc.
All of these factors would also need to be taken into account as well, including initial placement of Factions upon landing on Planet. Very difficult to catalogue.

4. The learning curve of humans: human player x plays one way in a game. Player x matures over time. Player x encounters same circumstances in a new game. Player x employs new, learned strategies, against the AI.

5. AI/ Human relationships: can an AI comprehend the difference between a human desperate for an AI's support in a joint attack, versus a human's intent of committing an AI's forces to destruction, thus bleeding dry an AI's reserve in order for the human to attack it personally?

While ideally I understand the logic behind the argument, I can say empirically speaking that there are simply too many variables to ever see this idea through to fruition. It would take years to build the database described, and in the meantime the AI's continueally uploading the latest "best guess scenarios" would too often missfire, ending in ruined games for a lot of human players involved. And by the time the database would reach fruition (i.e. where the law of averages would really begin to take hold) the game would probably be on the way to becoming obsolete, if not already obsolete.


D

That's why it's the hard way.

I still play the game on occasion against the current AI. Imagine how much better that experience would be if the programmers had tweaked the AI a little bit over time. I'm sure that some of us would be willing to fork over $10 or so in order to play against an AI which recognizes a few more tactics and strategies and can terraform better, ie things that modders are incapable of improving much because they are largely or entirely in the domain of the programmer. I can say confidently that I would have paid $10 a year in order to receive annual AI updates, the amount would have been trivial in comparison to the many hours of play that the game has given me.

As for the database of played games idea, the ideal would be a learning software that cranked away in the background analysing each player's game style (and altering tactics to meet it) with pre-chewed results being fed over the web to a larger database that the programmers could use to make larger changes to the AI algorithms. It is perhaps more science fiction than possibility currently, but it sure would be sweet.

Anyway, your points are quite valid and it certainly seems that simplifying the game is the option of choice for game companies. I hope that they do so with a lot more discretion than they did with Civ 3, which was dead on arrival as far as I was concerned. The simple truth is that in order to provide me with a complex and immersive environment that will keep me interested the game must be too complex for the AI to have a chance. The AI must play a different game than the human. It can work well playing a series of tactical mini-games against the human, but when it comes to complex systems like economics and diplomacy the AI is simply incapable of brute forcing its way into a serious competitor.

[binTravkin]

Or you can just make the AI capable of learning from mistakes in the first place.

As far as I know the universal algorithm of learning which works in a human mind has not yet been discovered for an AI.
Although we know the "see->save->see the same->use memory" sequence, computer just can't implement it in this simple way - it needs many things to be overlooked to even start saving things properly, not talking about using them on the correct circumstances.

We could, although tech AI some existing strategies and also remember good moves from the last games.
But the problem is that there are just too many variables to compare one game to other.
And eve if AI compares them both and say 99% of the variables match, the remaining 1% or even a single variable, no matter how much percent of the total it makes up can can change everything.

E.G. variable "enemy_has_nukes".

I'm sure that some of us would be willing to fork over $10 or so in order to play against an AI which recognizes a few more tactics and strategies and can terraform better, ie things that modders are incapable of improving much because they are largely or entirely in the domain of the programmer. I can say confidently that I would have paid $10 a year in order to receive annual AI updates, the amount would have been trivial in comparison to the many hours of play that the game has given me.

A good "heads up"!

As for the database of played games idea, the ideal would be a learning software that cranked away in the background analysing each player's game style (and altering tactics to meet it) with pre-chewed results being fed over the web to a larger database that the programmers could use to make larger changes to the AI algorithms. It is perhaps more science fiction than possibility currently, but it sure would be sweet.

And a brilliant idea about that database! It would be pretty hard to implement on a large scale, but such small tricks as "use AAA against chop&drop" or "bombard stacks" would help AI a lot together with the bonuses it already has!

[binTravkin]

The AI must play a different game than the human. It can work well playing a series of tactical mini-games against the human, but when it comes to complex systems like economics and diplomacy the AI is simply incapable of brute forcing its way into a serious competitor.

I think we could split up the programming AI in a couple of different modules:

1.Terraform
2.Governoring
3.Battle Tacs
4.Diplomacy
5.Global management (e.g. research, SE)
6.(reserved for something I've forgotten)

Some of those modules will have to interact, but the terraform one is almost a stand alone.
If we manage to get automation to a level it can create intelligent grids of terraforming, the AI will be so much stronger by that simple thing.

[Commy]

Since SMAC is infinitely replayable, perhaps the AI should be tought to learn (sentient econometrics, anyone?)...

Okay, so lets say that every game you ICS, so on the first game, you kill the AI, then the AI learns some tricks about ICSing, and ICS itself, then by the third game, the AI has learned how to defeat ICSing with a certain military strategy, so on games to come, it will always know how to defeat someone ICSing...

Or, perhaps in every game, you attack the Hive first, so one the first game, you attack the Hive, they die...second game, you attack, and they put up a good fight, but still loose...third game, they attack you first and kill you...

Of course, not every game then would be a completely new game, because whatever you did in past games affects outcomes in future games...perhaps a "AI memory" reset button would be needed...

[Senethro]

If you could invent an AI that could do that then you could buy Bill Gates.

How do you explain the concept of ICS to a CPU?

[binTravkin]

How do you explain the concept of ICS to a CPU?

Its easy.

But you can't explain the concept of "learning" to CPU.
At least giving the theoretical base we have nowadays - no chance.

[Senethro]

How would you tell the AI to spot ICS? Terraforming pattern? A certain ratio of bases/land area? Use of specialists and supply crawlers? I can think of plenty of exceptions.

[PJayTycy]

You can have a "learning" AI only on some domains the developpers specify when creating it. For example your ICS: the AI should notice the close spacing of the bases. So, the developpers should have a variable "base_spacing" which the AI observes and reacts to. If nobody at Firaxis ever thought about ICS as a strategy, they would most likely not have inserted the means to track that. Likewise for something that could be called "num_crawlers_working"...

An AI can only learn by adjusting preset variables. But, already with the example I gave above, the difficulties come clear. If you win with an ICS aproach, the AI will value close base spacing and not use a lot of crawlers. If you win with a builder aproach, the opposite will happen.

So, it becomes clear there should be another variable, that can decides when to use ICS and when to use builder strategies (and then a set of values for each of these 2 strategies). We can see this now, after having played the game a lot. Probably the AI developpers see it too after playing the game a lot. But, when you as a player notice this, it's a lot easier to just implement both strategies than to make an AI who will discover them by himself.

Also, the AI can play very good, but still lose. Then it "learns" it played bad, and adjust its tactics for the next game. However, now it'll play worse. This game has only 1 critic (did you win or loose?) and an almost endless amount of variables to be adjusted, while only 10% of them really mattered in the just-played game.

[Chaos Theory]

It's quite hard to write an effective AI for a difficult task like playing SMAC, but with the right interface, any fan would be able to give it a try. SMAC's AI is, as far as I know, tunable only through the meager mechanisms in the faction.txt and alpha.txt files, and sometimes the scenario editor.

Furthermore, it has been demonstrated that, with enough games, AI algorithms can tune many parameters with just the feedback of "win" or "lose". These games can be generated by playing the computer against itself. For more information, research Evolutionary Computation.

[paramir]

crawlers are most key

[GeoModder]

For a Morgan, sure it is.

[Kataphraktoi]

crawlers are just about the best unit for any faction.

[paramir]

[jk]says a morganite[/jk]

[Mead]

Re: How to improve the AI's?

Originally posted by Darsnan
So what would be the things you'd most like improved in the AI's?

To detail what I mean by each of the options above:

***

9. Use of Carriers/ Cloudbases: The Carrier is the most powerful unit in the history of naval warfare, and Cloudbases, even though featured in the CBA SP Movie, didn't make th final cut in SMAC(X). So, simply put, teach the AI's to use these powerful units once they become available!

***

D

D

I agree with just about all of your suggestions. It looks like a nice compiliation of the complaints many here have had about AI limitations.

Suggestion 9 got my attention. The few times I have made A/C carriers I lost them, with all the A/C on board, whenever they got in range (and detected) of the missile equipped AI. Even with a large number of SAM A/C included, and escorted by SAM well armoured ships I still lost my carrier task forces.

Mead

PS

I guess it would be more useful to only send them in range to do their strike then retire out of range, but even better was to rely on the normal get a foothold (via conquest, probe, or Needlejet CP new base) near the AI and launch my chop and drop from there.