[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.
The AI is dumb, let's face it. Even the best choice is often the wrong one, my testing proved that.
 
The AI is dumb, let's face it. Even the best choice is often the wrong one, my testing proved that.
That may be an overstatement. Are you sure that AI chooses wrong options more than good ones? How do you even measure that?
 
Can we try one version with randomization of two top choices, including social policies, for all difficulties?
You can try that right now by editing the Difficulty Options XML file.
 
edit: replying to cppmaster but I'm on mobile so everything hard

i guess i'm saying we have a slighly deceptive proposal here which might yield positive results only with additional efforts which nobody has committed to.
 
edit: replying to cppmaster but I'm on mobile so everything hard

i guess i'm saying we have a slighly deceptive proposal here which might yield positive results only with additional efforts which nobody has committed to.
I see, but also, as stated by Recursive, it would be easier to tweak scores when they are deterministic, so it might be a good direction.
 
Unless the scoring system is massively faulty (like giving huge score for warfare when playing an extremely peaceful civs) you can't really improve the scoring that much with a lot of variation between games and between balance patches. It's hard even for human player to really determine exactly if picking this choice is 5% or 10% better than that choice to adjust the score accordingly.

That's why it's better to just use a system that, while won't always pick the absolute perfect choice, guaranteed to pick only among the best choices available. Any big variation (like peaceful civs going full warmonger) should be treated as bug and isolated/tested/fixed individually and should not be seen as fault in the system that you need to tear it down and rescore everything again.
 
That may be an overstatement. Are you sure that AI chooses wrong options more than good ones? How do you even measure that?
It's not about choosing options, it's about diplomatic decisions, strategic insight, you'd have to reprogram the entire game from the ground up to account for everything the AI can do, even Vox Populi can't build a new AI system from the ground up.

You see an AI declaring war on someone and then getting overrun without the defender even having to make new units and you tell me the AI is smart.
 
I say no. Randomness contributes greatly to replayability. The last thing you want is two playthroughs feeling somewhat similar. Anything that contributes to this is a hard no from me.
Also, I thought it picks the top options more often the higher the game difficulty is set to. Is this no longer the case? Is there something wrong with this approach that it needs to be fixed?
 
Are pantheon/belief choices currently randomized? There's no column that controls that.

I'd agree on (unlimited) randomization on ALL difficulties only if the choices score 90+% of the top choice.

Example:
Choice A - 2000
Choice B - 1940
Choice C - 1930
Choice D - 1900
Choice E - 1830
Choice F - 1800
Choice G - 1400
Choice H - 1340
...

AI would pick randomly (weighted) between choices A to F.

But this will have to wait till next proposal session.
 
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.
edit: replying to cppmaster but I'm on mobile so everything hard

i guess i'm saying we have a slighly deceptive proposal here which might yield positive results only with additional efforts which nobody has committed to.
You make a fair point. It was not intentionally deceptive, but I should have edited my OP to reflect your clarification, that's something I forgot to do.

I still believe that the current system does not do the AI any favors, but I've edited the OP to explain more precisely why and I've reset the vote total for fairness.

As for additional efforts, I'll gladly commit to doing those.
 
Are pantheon/belief choices currently randomized? There's no column that controls that.

I'd agree on (unlimited) randomization on ALL difficulties only if the choices score 90+% of the top choice.

Example:
Choice A - 2000
Choice B - 1940
Choice C - 1930
Choice D - 1900
Choice E - 1830
Choice F - 1800
Choice G - 1400
Choice H - 1340
...

AI would pick randomly (weighted) between choices A to F.

But this will have to wait till next proposal session.
No.
 
I voted no again even after the clarification that recursive will actively try to improve AI scoring after randomization is removed, because in order to do so it's best to rely on data from lots of controlled test (same environment) thus best being left as an option through xml so anyone who want to contribute to the database can do it themselves. It might help making the AI better later with those data from removing randomization, but it's not something that should be pushed to general player base who just want to play a game with good AI.

But I'm not against making another proposal next season to add a certain threshold to limit AI from picking options that's too underweighted (like lower than 80% of the top score), which would make the AI better right away without a long period of testing > adjusting score > repeat.
 
Proposal vetoed.

Reason:
After considering @ilteroi, @azum4roll and @balparmak's points as well as how much work it would require to implement this proposal, I've decided it's not feasible to implement within a reasonable time frame. However, there is a clear consensus that players do not want to make the AI stupider based on difficulty, as both this poll and the previous one had a majority in favor of making this change.

Instead what I'm going to do is this on all difficulty levels:
CityNumOptionsConsidered: 90 (%)
TechNumOptionsConsidered: 90 (%)
PolicyNumOptionsConsidered: 90 (%)

Rather than pick using weighted randomization from the top X proposals, the AI will consider all options which have a score of at least 90% of the top score, and pick using weighted randomization from those options. This should adequately serve to prevent the AI from doing stupid things since an option which scored 90%+ of the top score should be good enough, allow for some randomization from game to game for the players who want that, and still allow players to perceive any policy trees which are chosen too often.

If anyone wants to turn this off (to experiment with AI scoring for debugging), they can set it to 100, which will only pick options that have the top score.

In addition, if anyone really wants to make the AI stupider, they can always set it to 50 or whatever.

I think this is a good compromise that can be done immediately to improve AI intelligence.
 
Status
Not open for further replies.
Top Bottom