AI is only forced to shop for currency and the troop upgrades.  FS1 has a high priority after war score 100 and is cheap, so is often purchased by the first AI to shop for a second time.  The score requirement basically removes it from the initial shop, so the AI does not spend their initial gold on that upgrade.  

Peppe, do you have spending caps on items for AI with open item slots?  If so, is it a percent, or a hard cap?

Most of the cause of early artifact buying has always been that the AI preferred the most expensive item it can afford, but I thought that was mostly taken care of in this version, until the AI has filled its slots - then it's just a matter of setting the right priorities for the right artifacts / expensive normal items.

Edit: Solution to the FS1 'problem':  Add an OR gold check.  WarScore >= 100 or Gold >= 8000 or whatever nightmare starts with.  That way nightmares will pick up FS1 immediately, without hurting their initial equip much at all.

FS1 is the only one with a forced delay to get it out of the initial shop list. The upgrade does not do much until buildings take some damage.

Miriyaka, Nightmare starts with 4k and should buy 2 1500 items and 1-2 other items.  Don't think FS1 should be in the initial shop.

The purchasing scheme I settled on is max item price for the first 20 seconds is 1500, unless it starts with more than 5000 (high gold or filthy rich).  If it has high gold or we are outside the initial shop the AI can purchase whatever it wants.     On a map like exile (artifact shop at start location) the AI will probably fill up on artifacts.

Basing it on open slots had it wasting money rather than getting the best item that could help it right now.

Nightmare on high has 30k. 

With the updated global priorities I did a test on cataract and generals got 5 purchases in the 20 seconds: Bishops/high priests, godplate, vlemish, journeyman, hauberk of life.

Assassins got 5 items in 20 seconds: godplate, vlemish, journeyman, nimoth, hauberk of life.   One got narmoth ring instead of vlemish (vlemish is zeroed in his build).

Based on current priorities nimoth(40)or HoL (40) will be replaced first by artificats.  Then journeyman (45), then vlemish(50) or godplate (50)

Excile start:

All fathers, vlemish, 1-2 cheap items (idols, banded etc)

Should I make some rule to handle rich starts?  It can afford god plate, but not all fathers.

26n:

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26n.zip

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26n.exe

Tweaked shopping for high gold start.

Global item priorities updated, godplate and journeyman purchasable now.

Moved pounce 1 to its own action.  The only trigger is interrupt.

Try some forced goal re-weighting to get the AI to do the right action after a teleport in to save a flag when it may not notice an enemy hero in its radius.

TY. I'll do another Sedna legion test with the same conditions a bit later.

Any silence could seem like a good time when there's 3 other Sedna's waiting with silence, how can you really tell if they're playing right when it's unlike anything you'd ever see in a real game. 

Hmm... Your point does make sense. So much is going on, that even a replay evaluation from observer POV can be somewhat problematic. Okay, I'm convinced: I'll do a balanced test. One AI Sedna on each side, and that's it.

double post

Falls behind what in level?  The enemy team?  Right now (if Peppe is using that xp priorityfunction I posted), if the AI stays lower level longer, it will be far more likely to get Exp 1 and 2.  But no comparisons are done to the enemy team, and it uses the team average level when deciding on the upgrade's priority, not its own (for obvious reasons).  However, there is also a time cutoff, after which it will not buy Exp 2+ at all.  The low-XP game setting doubles the length of the time cutoff as well as the base priority, I think.

Re: Reg snipe build, I assume it's using Staff of Renewal?  I know there was some back and forth about deciding between SoR and BotS, and I just do not see BotS being worthwhile with a snipe/mine build.  The AI is not smart enough (and probably will not ever be smart enough) to use BotS at just the right time to recover mana from snipes (snipe works with BotS, yes?) and from concentrated creep AA farming, and it's better to give it the ability to snipe more rapidly.

So - how much of a pain in the ass would it be to change ai logic some based on its level (eg, if its normal it purchases this cit upgrade... if its nightmare, it purchases this)?

Another thing.  I'm not sure if this will hurt things on normal, but after playing a couple of games with cow on nightmare with 1 ai, I really think the ai should buy cur2 asap as well (ws5).  In many MP games, getting cur2 is a bad idea; however, with an ai on your team, I'd rather he take the hit instead of saving for this or that.  This was more clear when our ai reg with 11,000 gold choose to buy godplate over cur2. 

I guess that fits my underlying philosophy with the ai mod.  My #1 goal is to improve the experience of players (so an AI gimping itself to help them team is cool with me).  My #2 goal is to have the ai play competently.

Oh - your logic change so that the ai buys jtreads is great.  The ai manages to avoid alot of deaths with this item and it very much compensates for any weak logic.  Its a good buffer for things we haven't tweaked.... granted, its primarily viable in nightmare AI games, but still - its a good tweak.

I wonder what we can do to make the ai even more ridiculous late game though.  Us experienced MP folks adapt fairly quickly to the new AI.  We know that it will be very strong early game (stronger than we could be), so we bide our time.  Come mid-late game, we can crush it.  We need to figure out how to change that so its difficult the entire time. 

I think we could cover that goal with a few logic changes.  For instance, perhaps 1 AI could be designated as OP.  That AI would only purchase items for itself and bypass citadel upgrades.  Having 1 super strong opponent would be helpful, but I'm not sure if you can accomplish that with coding.  I'd actually say - we provide a list of specific dgs that should be considered for OP status.  We'd then provide a specific item build for the OP dg and see how it does.  The downside is that this fits much better in a nightmare level AI game and wouldn't be all that super in a normal AI game (which takes me back to the first sentence of this post).  Anyway, just some thoughts. 

Extremely trivial, in fact I think I already have at least one function (currency) doing just that.  I'm starting to talk out my ass here, and should probably look at what Peppe actually has in the mod right now, but I did add a difficult mult function into AIGlobals.

Edit: From my comments in AIGlobals:

Currency II: 8/12 minute cutoff for normal AI on normal/high gold, 16/24 min for nightmare AI, starting pri 40

Currency III: 6/9 min for normal, 12/18 for nightmare, starting pri 20

Looks like Peppe halved the priorities though, 20 and 10 times gold mult respectively (high gold is x1.5).  So cut those times in half as well.  Was normal AI actually buying Cur2 that often?

We can up the priority on currency 2 and 3.  They are not great upgrades though even with the nightmare gold bonus.  For currency 2 the AI that buys either one isn't going to recoup the costs by the end of the game.  The team as a whole may come out a little ahead in 4-5 minutes.   Currency 3 is hard to imagine a setup where this pays off.

On ashkandor we can force the AI to save 18000 for any artifact purchase, then when it goes to shop if ashkandor is in its list as the highest priority it will purchase it.  Otherwise it will probably buy bulwark.  In that situation it is only over saving 2k.   With all the upgrades that are set as purchasable this is not something you will see in normal games as even a team of 3 normals spend nearly all their gold on upgrades.   Hard and nightmare have a small window before the armor/damage upgrades become a priority to spend some cash on artifacts.

Just use a PriorityFunction to put Ashkandor at higher priority if IsAssassin(unit)?  Sure, they might shop between 16k and 18k, but maybe half the time at most.  Only Nightmare will be affording it anyway under normal conditions (not high gold).  It feels like item exceptions (other than the starting cap) will be something that will make it that much more difficult to tweak item priorities in the future, without providing that much more benefit.

Looked at Currency functions again, and I had things backwards.  Nightmare was actually less likely to buy than normal.  Oops.

Try this instead: http://pastebin.com/Ub9qb7b9

Made it so nightmare gets double priority, time cutoffs for II are now 7/11 normal, 15/22 nightmare; III is now 5/7 normal, 10/15 nightmare.  Keep in mind that priority decreases over time, so even within the time threshold, nightmare AI may still buy something else instead as the priority slowly drops to 0.

Experience had the same problem in the Skirmish AI copy, but I already had my local copy changed so that difficulty increases base priority.  Cutoff for Exp II for normal AI is average level 10 @ 10 minutes, with time and level weighted equally (e.g. 10+10 = 8+12) and a base priority of ~50.  By the time it's available, base priority should already be below 25-30 for normal AI.  No AI should ever buy IV except maybe Nightmare on low-XP games, but even that is very very unlikely.  Normal should never buy III.

Exp functions: http://pastebin.com/xUsTrEsN

Keep in mind the constantly reducing priorities for these upgrades when you go to adjust them.  Both Exp and Cur II should have somewhere around half their base priority for a normal AI by the time they become available in a normal WR1-start game, which should put them similar to or below most items.  Also make sure you adjust the multiplier for time/level proportionately so that it doesn't radically lower the cutoff time.