AI shuts down ALL research when it reaches 'Zero Gravity Constr.'

This...doesn't sound good. I'm curious now, too....

I noticed in my first 1.2 game that Zero Gravity Construction now takes a lot longer to research than it did before 1.2.  Perhaps the AI is looking at the research cost and thinking that it's not worth it to research?

That might explain why in my latest game two of the AIs just stagnated and appeared to stop researching all together. Although that didnt affect all the AIs, just 2 out of 5.

They increased the base cost of Zero Gravity Construction from 88 points to 512 in the last patch. This is further compounded by research inflation, which is effected by both a multiplier and an exponent. The exponent is 1.2, and the multiplier is a 12% increase. I should note, I'm not 100% sure of the exact order those are applied in; but those are the modifiers listed in the files, and either way you get similar results to the below.

So let's say you research just 1 tech prior to ZGC. ZGC now costs (512^1.2)*1.12, or 1995.84 research points. That's a 4x increase from having 1 tech. Once you have 3, it's costing tens of thousands. 

Now, the AI seems to get all tech costs halved on all difficulties. But it's also horrible at production, and may only be spending 30% of it's budget on research, much of which is wasted on planets with no research buildings. It's very unusual for any AI to ever go over 1000 research output per turn below Godlike, and even then it takes a hundred or more turns.

So, if it's got more than, say, ten techs before it attempts ZGC, it may simply be taking the AI empires which pick it later on several hundred turns to learn the technology. Large-scale construction now has a base cost of 1078... which swells to nearly 5k if you've developed 1 other tech first. And it's half a dozen techs down the tree.

I'm not entirely sure what SD were thinking when they boosted these tech costs; possibly they assumed that everyone rushes them so the values aren't going to be too high by that point. But the AI blunders into them much later, when they're fantastically expensive, and so is utterly incapable of ever amassing enough research to pick them up.

I noticed that they had increased the research cost of some techs and I assumed they thought that slowing the human player from grabbing the larger ships early would help the game. I haven't yet seen the long term effect on the AI but if it is as it appears, you have to wonder if they do any play testing.

Maybe they should enlist some of the better players, (not me, by any means), to try out some of theses things before putting them in. If this is as you have described, it seems to be serious breakage.

I'm fairly sure that's what the opt-in is for tbh. Problem is, the final opt-in it only goes onto Steam a day or so before it's actually due to be released, and often the release version is exactly the same as the opt-in from only a day or two before. The monthly schedule of updates is great... but doesn't leave much time for changes to be bedded in before they have to be released to the public. Much the same thing happened with the BestDefence fiasco with 1.1; they added that change in very late in the opt-in, and the patch went live only a day or two later. The opt-in is good for spotting technical problems, but balance stuff takes several games to spot and there's not really time to start pattern recognition in that time period.

If that is the case, WOW, the AI is even worse than I thought it was.  I've read your posts enough to know it is very likely you have a sound basis, but as such an extraordinary statement, could you let us know more?

They SERIOUSLY incresed the cost of large hulls as well. They also did a number on Beam weapon costs post Disrupters. 

From galciv3AIdefs:

It's under the difficulty levels, but isn't part of them. This is not overly surprising; not many AIs could cope with the exponential cost increase, and technically speaking there is no strategic AI. It's all scripted. It's some dude at SD taking a guess several days ago at what would be the right global production settings 50 turns into the game you're starting tomorrow. It is not related to the events in the game. It does not think 'this is a killer tech, I should up my research output'. If you're very lucky, you can script it to probably be concentrating on research at about the right time in the average game where this cool tech might be one of the options about half the time. And this gets worse the later in the game it is. I can script a monster AI for the opening 50 turns. The next 50, not so much; the 50 after that, even worse; by turn 200 or so it's dead on the vine.

Being able to specialize, the human's research output increases by hundreds at a time, and so he can play the Red Queen on the exponent - running as fast as he can to stay where he is. But the AI's output increases in ones at twos. Even a strong AI would likely struggle with exponential cost increases anyway - it's brutal, and actually forces the player to specialize - but without specialization as an option, it doesn't stand a chance. Even Godilke AIs, with their vast bonuses, fall well behind the tech exponent by the late game if you remove the tech cost divisor.

I suspect the economy is presently extremely lenient for similar reasons. This is largely why I say that the AI colony specialization is one of the most important things that needs to be added. Until it can do that, several of the major balance issues in the game cannot be fixed without completely crippling the computer players.

Oh, and it should be noted that in 1.2 they added a modifier that means AI ships don't pay maintenance if they're in orbit of a planet or SB. This was added to all difficulty levels, even Easy. I also suspect that the -100% maintenance modifier stacks with other negative ship cost multipliers, so that the AI can actually earn money from defense ships if it picks up maintenance reductions, but this is hard to verify in-game. The code structure certainly implies it, though.

I'd be surprised if the AI was getting a bonus on research like that (as an XML).  I THINK it has to do with evaluating what tech it should research by dividing the cost of the tech in half when it does its calculations on valuing it.  

Let me put it another way, if we were going to have the AI get a research bonus, we'd just do it in code. It makes no sense to have an XML line like that.

Re the researching on those particular techs, I"ll ask them to look into it.

Is this really true? I remember Frogboy repeatedly posting that there is no scripted AI, but all decision are made based on calculations, asserting values to options and weighing them?

I don't think the AI actually checks the costs at all. Everything else in the xml implies a (fairly complex) triple-weighted random array. And I'm not sure I see how a universal halving of all values would make any decision on such an array give a different result than without it.

But it should be easy enough to check - if we just set the divisor to something silly-high (like 20) and run the game for a hundred turns, then see if the AI has picked up any more tech than usual.

That's by the by, though; if the AI doesn't have it's tech costs reduced in any way, then it will be even more brutally crushed by the exponent and the giant increase in ship tech costs. 

Oh, and incidentally - a generic tech cost decreasing modifier in a stats field would be a much better way of handling research speed settings in the game setup. The present bonuses in there stack with buildings, and so become less effective the more research buildings the planet has. Very Fast research is basically just a bonus T4 lab per planet presently.

It's true, though with qualifiers - the actual AI part of the AI is a data-driven AI engine. This handles stuff like 'where will I send this ship?','should I risk attacking that fleet?' and 'that fleet has lots of missiles, should I send my PD-heavy fleet to get it?'. But large chunks of the computer player's strategic behaviour is heavily scripted; it's build choices (both buildings and scripts), it's ship designs, it's economy settings; how many ships to station as defenders.

The AI part of the AI is very good. The scripted part is not. Players presently are largely running rings around the scripted bit, and so presume the AI is bad. It's not. The AI is trying to do a good job, but is relying on being given the resources to do it's job by the brain-dead scripted bit which simply cannot provide what it needs; it's like having Napoleon leading the army but relying on an economy that's being run by Robert Mugabe to support it. 

Thanks for explaining, very helpful! 

So I was looking into some of the issues mentioned above, good stuff in general, and I have already been able to address a few issues. Mostly cost balancing of some techs. These changes will be in 1.2.1 next week some time.

I did look into the ResearchCostDivisor from the AI Strategy defs, and in fact it does not reduce cost of techs for the ai, it is used as part of the weighting calculation when the AI is deciding what techs it wants, and how valuable they are. We have to have it go up in a curve or it tends to just go straight up a branch to its end. This Divisor is simply there to set the slope of the curve.

Thanks for the feedback, it is always welcome.

That's good to know, thank you