DESIGN: AI Improvement discussion

[Tamerlin]

You all know my point of view, the less an AI cheats the better as far as I am concerned. I would prefer the AI to be taught how to manage its resources and respect the rules rather than giving it another advantage.

I don't like games in which only the human player has rules to obey to, if the rules are not the same for the AI where is the strategical part of the game?

[MrBaggins]

I'm all for as naturally a tough AI, without as many artificial benefits as possible.

Having said that, I'm also after a challenging and competitive game. Just as long as the cheats aren't "in my face".

A simple fact, Tamerlin, is that the state of AI is really pretty crap. In a game like CtP2, you are trying to emulate human behavior. The technology just isn't there to do that. There is no simple way to simulate learning: connective-reason based adaptation.

Without that, an AI will always be second class. You can try to think of all of the conditions and how to deal with them, but a human will always "plan b" you.

So... you either put up with some amount of AI bonus, and foreknowledge... or you have a ridiculously easy game.

Which would you prefer, Tamerlin?

[Gilgamensch]

Actually a similar system as I mentioned above (table) could be used for the AI to optimise cities.

this table would kind of look like:

74,0,0,0,0
73,0,0,0,0
72,1,0,0,0
71,1,0,0,0

First digit being hapiness, second # of entertainers, third # of workers...etc.......

We could include a modifier in the end (like a counter) to sum up when it shall change (i.e. 3 for -69, if it would have 20 points, time to change the national settings.....)


Only an

[MrBaggins]

Hmmm.. hadn't got there... and its a bit more complicated than that, since you can consider whether its better to change empire sliders globally or build happiness improvements instead. Then you get into considering Guns or Butter, and how that should be prioritized.

But, yes, its ridiculous that the AI doesn't ALWAYS use entertainers to keep cities out of rioting.

[Gilgamensch]

Actually those digits you might first run through the 'city manager' to decide upon building hapiness buildings (including polution).

I think it is just a question of changing the code and creating 'tables'.

But again haven't checked the code yet.

For sliders: Normally you use them once you run out of buildings/entertainers as well as in this order.

Or if you use in general to much entertainers..........

By the way, sometimes also depends on the city size..........

[Tamerlin]

Originally posted by MrBaggins
So... you either put up with some amount of AI bonus, and foreknowledge... or you have a ridiculously easy game.

Which would you prefer, Tamerlin?

I agree with you MrBaggins, but I would just like the cheats to be limited to the minimum. If an AI cheats I don't want it to be blatant (as in Civ2 or Civ3) or/and crude (as in Civ3) thus spoiling the pleasure I can have to play the said game. Giving some advantages to the AI is the most discreet one and does not worry me.

[Solver]

AI cheating should be as transparent as possible. That's the point, generally, not how much it cheats.

For instance, the CtP2 AI cheats blatantly with city sizes and expansion. That's very visible and spoils it for me. In Civ 3, if the AI only gets a 10% production boost, it's not too visible, but enough to make the AI stronger.

[Peter Triggs]

I think that the blatantly obvious AI cheats are own fault. In order to try to make the game more challenging, most mods tweak the data in DiffDB to the point where the AI's advantages can become too obvious.

[MrBaggins]

Simple and cheap method for threat determination

So.. in a moment of brainstorming, I was looking to find a cheaper method for measuring the threat to a specific map point.

You can use threat mapping, but it can tend to get computationally expensive, so I considered a different data model, which should scale well, if be initially less precise;

A "radar" system

an area which the AI wishes to defend... a city, say, has a (complicated by the tiled map location system) x, y location.

You can iterate through all the armies to find units within a particular range, calculated from the x,y, not dealing with the actual travel distance yet.

You then have a quick data set, which you can use to determine the total volume of possible threats.

Where appropriate (basically where the threat was significant, and where there were insufficient local defenders to deal with the problem,) more detail can be processed about the "radar picture"... actual travel time, troop types & strengths.

Its actually better than an influence map: While it wouldn't accurately describe the threat at a distance (it doesn't understand direction... just range,) its actually more accurate, because it deals with flanking; it would "group" two discrete threats approaching from different directions, where an influence map wouldn't until they were right on you... the map wouldn't describe the actual threat combined until distance was 0, which is too late.

Incidentally, cities, or other map points, within the context of the AI should include a data structure to include spillover defensive troops, so that theoretically more than 12 troops can defend a location, 12 in the city and more loitering outside.

[hexagonian]

Originally posted by Peter Triggs
I think that the blatantly obvious AI cheats are own fault. In order to try to make the game more challenging, most mods tweak the data in DiffDB to the point where the AI's advantages can become too obvious.

The real culprit in the Mods is the PW/gold bonus SLIC file. I drastically scaled down the PW/Gold SLIC cheat for my personal Cradle setup - the end result was that it was a lot more subtle and realistic. Most cities were about the same size as mine, and the amount of tile improvements around them were only slightly greater than mine.

And the tech gap certainly wasn't as drastic either.

As I noted in the city cap thread, most of the civs were close to their caps too. Once I was able to get the civs to do that, the inflated bonuses (especially from that SLIC file) are not needed.

[MrBaggins]

Originally posted by hexagonian

The real culprit in the Mods is the PW/gold bonus SLIC file. I drastically scaled down the PW/Gold SLIC cheat for my personal Cradle setup - the end result was that it was a lot more subtle and realistic. Most cities were about the same size as mine, and the amount of tile improvements around them were only slightly greater than mine.

And the tech gap certainly wasn't as drastic either.

As I noted in the city cap thread, most of the civs were close to their caps too. Once I was able to get the civs to do that, the inflated bonuses (especially from that SLIC file) are not needed.

What are the AI bonuses you're running with in diffdb, same as standard Cradle?

[Peter Triggs]

The real culprit in the Mods is the PW/gold bonus SLIC file. I drastically scaled down the PW/Gold SLIC cheat for my personal Cradle setup

I did it a different way: I didn't use that file at all. Instead, I put in an infrastructure workaround which simulated the infrastructure thing in the AI's buildlists (which it never used). Instead of churning out unnecessary and support costly units, I could have the AI switch some of it's cities to infrastructure (like I, as a human player, do).

BTW, I have it set up so that this option becomes available with Slave Labor, i.e. right from the beginning. CTP2's public works system is just so much better than civ2's. The other thing they put in, infrastructure corresponding to capitalization, was another great idea but they didn't exploit it enough.

[MrBaggins]

Just a little note on this. I was reading (in B&N) a couple of what are apparently the best conceptual books on AI coding...

AI Game Programming Wisdom (1) and 2...

I've ordered them from Amazon, and they'll be arriving directly.

They demonstrate an approach to deal with transport of units (and cooperation of units in other ways) by using FSM's (Finite State Machines, not Fuzzy State Machines, which would be FuSM's.)

Individual armies each are their own FSM, communicate with others, using set rules, and goals, act in a predetermined although sensible and useful manner. This allows cooperation, and realistic tactics, such as naval invasion, combined attacks and defense.

FSM's are typically used in FPS's and RTS's and have shown a lot of promise in teamed behavior.

The FSM approach would work well for tactical movement, but we'd probably want other approaches for decision making elsewhere.

[child of Thor]

Sounds intresting, the only negative thought i could come up with is that maybe the amount of 'book keeping' needed in a game like CTP2, as compared to an FPS. We can get a lot of armies and individual units in these games, more than you would get in a normal non internet connected FPS game?

He he - have fun with your book, i'm still trying to plug my way though 'Code Complete' and a book on C++, its hard going

[MrBaggins]

Yes.. FPS's have far less actors, but they act in real time, and FSM's are really quick, and ultimately the least part of the processing taking place, anyway... very little processing is done per communication, or query. Most of the CPU work is dealing with various DirectX processes.

RTS' also use FSM, in real-time, but do those with far more actors (and every unit is individual, rather than in the civ concept, where you have groupings.) These are more akin to the amount of work necessary, and if it can be done in real time, then it definitely can be done in TBS.