Statistics about AI and Player spending

[essmene]

Hi,

having let the AI play a few turns in a few games (and totally messing up any tech, money, ... lead) i was wondering if anybody pulled statistics for

• spent on / / / Units / Buildings
• spent on / /
• population spent on Draft / whip
• spent on hurry buildings / unit upgrades

for a human player and for the AI. E.g. yesterday i had a cultural win in 30 rounds and set the AI to auto play.

Not only did it manage to spent my 5000 gold in a turn (while i had the power lead) - it also messed with my third cultural city trying to turn it into a unit production city (with 4 cathedrals / Hollywood,.... inside)...

For the time being i belief the AI 'looses' a lot of gold by upgrading old units where it might be 'cheaper' to just produce new units and abandon old ones.

At the moment i belief the AI spends like 50% of it's wealth on upgrades where i would assume i would spent about 10% for of it for my most experience units (and having less money than the AI).

 

[Minor Annoyance]

Well it might be because the AI only took over at a certain point, that it didn't have it's strategy set to cultural the way it was "set" for you. It may not have known the right way to pick up the game from where you left off or hadn't reevaluated it's strategy yet and was using a default starting strategy. A near cultural victory wouldn't sneak up on it if it was playing the whole game itself.

 

[essmene]

I would have hoped it would evaluate the winning conditions.

Back to the statistics. I think it would be good to know how, where the AI does spent it's gold / ressources on and if there could be a way to optimize it. E.g. not upgrading Swordman -> Rifleman, but simply deleting the unit might save a few thousand of gold for research.

 

[Roland Johansen]

Upgrading troops is almost always efficient for the AI as it only costs half the normal human player cost to upgrade units for the AI.

 

[jdog5000]

Since cultural victory requires a long term commitment, the AI decides whether or not it's going to try basically from the start of the game. If the personality and random factors do not have a high enough value at the beginning of the game, then it will not try. When the AI took over from you, it is unlikely to have trigger its cultural victory logic.

Going for victory is one of the things I am working on, right now the only one the AI tries for coherently is cultural victory. My plan is to make it so that in the early game the AI picks a potential victory target based on personality, game state, and random factors. Then, as the turns go by, it keeps evaluating which victory condition it is getting closest to and will pursue the best course.

There are a few other reasons that the behavior of the AI is far from its normal actions if you turn over control using autoplay late in the game. When the AI builds a rifleman, it assigns it a particular UNITAI, meaning this rifleman is intended for offensive operations or just as a stay at home defense man, etc. When you've built the rifles it inherits, then I think they all get the default UNITAI type for that unit. Such a situation, where the AI has plenty of units but they're not assigned in the right distribution to the various unit tasks, doesn't happen in the course of a normal game and so there is no mechanism for intelligently handling it ... instead, the AI will usually drop everything to fill what it feels are the big holes in its army makeup.

Letting the AI play most/all of a game on autoplay to watch what it does gives a fair assessment of its actions as it plays exactly the same as it would have if it was autonomous. Turning over control though produces abnormal behavior that can be look clearly suboptimal.

As for your question on statistics, no we don't have anything like that. I was planning on running a thread to evaluate the first 100 turns of the AI or something like that, getting community input on whether the early game decisions the AI made were good or how they could be improved. We'd probably learn a little bit about those kinds of details from such an experiment.

 

[TheMeInTeam]

As for your question on statistics, no we don't have anything like that. I was planning on running a thread to evaluate the first 100 turns of the AI or something like that, getting community input on whether the early game decisions the AI made were good or how they could be improved. We'd probably learn a little bit about those kinds of details from such an experiment.

I think we'd learn a lot for both the AI and ourselves . I'm sure some player's opening 100 turns vary wildly between games, but are there constant factors? Priorities? Patterns in tile improvement distribution? CONDITIONAL (but consistent within those conditions) choices? I'm curious just to improve my own play but if there are aspects of elite players that could be consistently replicated without hurting the AI performance (IE certain behaviors are consistently good) it could improve the AI a lot too.

 

[jdog5000]

Well, with the AI we'd first be looking for clearly bad choices ... then move on to see if there are consistently good choices they could have made.

But yeah, it would also be a great learning experience for all involved for sure.

 

[essmene]

I was hoping it would burn down to some Pareto Principle. 20% of the errors cause 80% of the damage.

The problem is what the AI is supposed to do. Sometimes i capture enemy capitals with like 3-4 tiles improved land, where my own has ~12+ around 0 AD. But on the other hand i build a lot workers, while the war lords do use their hammers on armies.

Another thing is what is an intrinsic problem with the current AI and what can be solved with scripts. I have read that city specialization is not possible with the AI, but i find it vital having 2-3 (or ~ 1/10th(?)) of my cities as army producing ones, one wonder spammer, one beaker, one commerce.

Looking at the first 100 turns sounds great with a list of the 3 worst things to avoid - and maybe the 3 best things it should have done.