[Vote] (1-15) (VETOED) Proposal: Remove Difficulty Randomization of AI Choices

Approval Vote for Proposal #15 (instructions below)


  • Total voters
    44
  • Poll closed .
Status
Not open for further replies.
Fair enough. Like I said, for the recent analytics projects we've been running on AI performance, having a mode where they aren't accidentally ruining themselves is ideal.
 
No. It's better to get $10 instead of $1 100% of times, than 90% of times.
Gameplay variety aside, that's only true if you assume the scoring system is perfect. In case of top choice at 100 and 2nd choice at 99 due to a slight mis-scoring, weighted random choice would still have 49.5% chance of taking the actual better option, rather than a 100% chance of picking the worse option.
Secondly it makes AI very predictable/railroaded into the same pattern, thus it's actually harder to judge if the scoring is good enough (since you can only see 1 out come every time in the same situation). If there're multiple outcomes you can at least deduce how the AI scored specific options based on their weight, and you can match it with your own experience to know if that scoring is good enough or not to adjust accordingly. Having AI only pick the top options you will never know if the 2nd choice is 99 vs 100 or 10 vs 100.
And lastly, gameplay variety.
 
Should we be instead adding randomized choices to deity AI?
I think this would fun to try, at least briefly. Partially because the scoring system isn't perfect, but also because it would cause variety.

A good example is that Oracle almost always gets built before Parthenon on Deity. This makes the parthenon a very safe choice for the player.

It also means there are AI who waste hammers not getting the Oracle. There is easily a situation where a random fluctuation to Drama before Philosophy not only hurts the player, but could help the AI, because now they get a wonder, where as before they would have missed it.
 
Gameplay variety aside, that's only true if you assume the scoring system is perfect. In case of top choice at 100 and 2nd choice at 99 due to a slight mis-scoring, weighted random choice would still have 49.5% chance of taking the actual better option, rather than a 100% chance of picking the worse option.
Secondly it makes AI very predictable/railroaded into the same pattern, thus it's actually harder to judge if the scoring is good enough (since you can only see 1 out come every time in the same situation). If there're multiple outcomes you can at least deduce how the AI scored specific options based on their weight, and you can match it with your own experience to know if that scoring is good enough or not to adjust accordingly. Having AI only pick the top options you will never know if the 2nd choice is 99 vs 100 or 10 vs 100.
And lastly, gameplay variety.
Ok, that makes sense. Still how could the scoring system be improved if it's flawed?
 
It also means there are AI who waste hammers not getting the Oracle. There is easily a situation where a random fluctuation to Drama before Philosophy not only hurts the player, but could help the AI, because now they get a wonder, where as before they would have missed it.
So it's like AI would make a good choice by an accident. If getting the Oracle is that good than maybe AI scores this tech too low.
 
Ok, that makes sense. Still how could the scoring system be improved if it's flawed?
You don't need to change the whole system, just adjust the scores for each elements that contributed to the score of each options. Balance patches happen frequently, thus those scores should also get adjust accordingly. The system itself is working fine.
 
You don't need to change the whole system, just adjust the scores for each elements that contributed to the score of each options. Balance patches happen frequently, thus those scores should also get adjust accordingly. The system itself is working fine.
Yeah, I didn't mean to change the whole system, only tweak it, so it's less flawed.
 
I think this would fun to try, at least briefly. Partially because the scoring system isn't perfect, but also because it would cause variety.

A good example is that Oracle almost always gets built before Parthenon on Deity. This makes the parthenon a very safe choice for the player.

It also means there are AI who waste hammers not getting the Oracle. There is easily a situation where a random fluctuation to Drama before Philosophy not only hurts the player, but could help the AI, because now they get a wonder, where as before they would have missed it.

There is certainly a bit of an issue with them being predictable. There are certain techs they love and always seem to take first out of that era. Like Railroad is fine but they don't all need to take it first every time.

This does flutuate a lot between patches too to be fair. It used to be possible to get terracotta army 100% of the time without that much effort because the AI hated it
 
I didn't know it was weighted either, but my vote doesn't change. If there was an option to have weighted randomization with a maximum score difference limit (let's say 20% of the top score), then I could be willing to accept randomization.

E.g. randomization between 500 and 400 is acceptable, 500 and 100 is not (even when weighted). The former is a choice between two good options, we can't be sure which is actually better since AI coding isn't and can't be perfect.

The latter on the other hand is a choice between one good and one bad option, it's not taking chances or experimenting, it's willingly sabotaging yourself at this point. If there's a case where 100 is the better choice, then scoring is very wrong and needs to be fixed. And we'd get much more consistent reports about this when the 100 option cannot ever be chosen.

I'm still on Yea but will ask the devs about feasibility of a max. score difference limit, may propose it at the next session.
 
Last edited:
So it's like AI would make a good choice by an accident. If getting the Oracle is that good than maybe AI scores this tech too low.
No, its because the more AI choose tech X, the better tech Y becomes. So by forcing all the AI to always do the same thing, an imbalance is created. On Deity it isn't uncommon for 7 AI to enter industrial era through the exact same tech, its very exploitable. I'm very confident that more randomization would make Deity harder, not easier.

There was the very good example posted before where sometimes they will assess two things, thing A scores 600, thing B scores 595. Should they always choose thing A? If people are really worried about the AI taking odd social policies, what if we just add randomization for everything but social policies?
 
Do AI traits factor into score-weighting? If so, could part of the variety/variance problem be that leaders aren't respecting their personal traits when scoring what they "think is best"?
 
No, its because the more AI choose tech X, the better tech Y becomes. So by forcing all the AI to always do the same thing, an imbalance is created. On Deity it isn't uncommon for 7 AI to enter industrial era through the exact same tech, its very exploitable. I'm very confident that more randomization would make Deity harder, not easier.

There was the very good example posted before where sometimes they will assess two things, thing A scores 600, thing B scores 595. Should they always choose thing A? If people are really worried about the AI taking odd social policies, what if we just add randomization for everything but social policies?

This is more of an issue with that particular tech's scoring then the lack of randomization. Scoring is/should be influenced by policies, pursued victory type, leader flavors etc. When taken together it should be very rare for 7 (I assume on the Standard size, so all?) AIs to have the same tech as the top scoring option. If that happens, that's an issue which we realize thanks to the lack of randomization and needs to be addressed.

Again, it's a misconception that the lack of randomization will railroad all AIs to choose the same thing, top scoring option for different AIs will/should be different depending on AIs status, needs and policies even without randomization.
 
This is more of an issue with that particular tech's scoring then the lack of randomization. Scoring is/should be influenced by policies, pursued victory type, leader flavors etc. When taken together it should be very rare for 7 (I assume on the Standard size, so all?) AIs to have the same tech as the top scoring option. If that happens, that's an issue which we realize thanks to the lack of randomization and needs to be addressed.

Again, it's a misconception that the lack of randomization will railroad all AIs to choose the same thing, top scoring option for different AIs will/should be different depending on AIs status, needs and policies even without randomization.
But it can be an issue with lack of randomization though.
For example at the start there might be some big score variation between choices, but because of the lack of randomization most AI picked the highest score options. Then the option they picked earlier can influence the next choice (to make sure they don't simply jump all over the place with no consistency) thus increasing the score for going into the same direction with other AIs who also picked the same 1st choice, and the process repeats itself until they got into a folk where the next set of options doesn't get influenced (or less) by all previous choices where the first 2 top scores are close enough (based on individual AI situation). If, even here, AIs still keep picking the only best choice, then we will get the whole process repeated again, leading to the issue CrazyG explained.
Having weighted randomization would certainly help at those specific folks.
 
dang, should have vetoed this. i kind of assumed you had given up on it :)

always picking the "top" choice even when there are multiple option with a similar score makes the AI worse. not better.
Yes, in some cases - but it allows us to observe the problems and improve on them.
 
Last edited:
I still think this conversation is trying to solve two separate issues at once. Firstly, that deterministic choices allow for better vetting of AI decision making, scoring, and allow for improvements in both realms to enhance the mod. Secondly, that games are interesting because of variance in AI choices. Maybe it's worthwhile discussing the merits of each approach, but at a certain point they're both valid for different purposes.
 
Last edited:
I would say that variance from playing against different civs under different circumstances is enough to make variance to the scoring and thus decision making.
 
I voted yes but only on the basis that I anticipate future improvements in the scoring based on feedback. Ever the optimist!

I've been playing for a while now with a tweak to the choice of what to build in cities, that if the first choice score is at least double the second choice it always picks the first choice, otherwise in proportion to the scores as before. If the AI really needs something, it can't be thwarted by RNG, but when the choice is close(ish), there is still some randomness.

Also, some situations, like what to build, often have lots of choices available. For others, like picking policies or tech, there are usually only a small number of choices. Perhaps different rules should apply in different cases.
 
so who will implement all those magical scoring improvements which likely will have side effects?
 
Status
Not open for further replies.
Top Bottom