Thanks.  A second set of eyes on the code is great.

I don't recall copying the citadel functions, so if its duplicated its probably just a carry through.  I'll take a look.

I'll see if i can cleanup the capture function to get the flag unit rather than search for the closest flag.

Retreat is mostly controlled by unit sensor.   There is a health sensor in there with a sliding 'oh shit' scale.  When that gets high then it goes and finds the best action to increase health and survival.  You have to look at action weights or goal sets to see what actions are likely to be called.   I think in that sensor file there is also an on take damage function.  You might be able to use that in other places to get a damage rate.

We have gone back and forth on letting the AI teleport in combat vs not.  It is in the teleport to health statue action. By default there is a distance check for enemies that blocks the action.   If i recall it kicks it to a find safe position.  Once it gets to a safe spot it uses the TP, but it may never make it to the safe spot.  Potions had the same enemy check but i took it off as most situation can get off a potion.

I don't remember if i updated them all, but Demigods that have an interrupt skill have some logic in there to interrupt their target when it detects an action started.  For sure UB's grasp is setup for that.

Found the other spot the AI interacts with its minions.  On teleport when it completes all minions are told to guard their master.  Going to see if I can count the healers and send some some back out on teleport.   Would also like to get it to cast higher level healers when it has them right now it keeps its existing healers until one dies and it will re-cast the higher healer.  Think I may do a release of what I have...

Looked through unit sensors and mostly understand how it works, but I'm not really sure what SensorStatus compares against when deciding new goals?  I tried to follow the chain that this value takes, and sort of lost it around recalculating goal insistence or some such.

Also, the value that the OnTakeDamage callback is adding to / subtracting from is probably being totally overridden by the HealthSensor levels, which just hard-set it and don't check the existing value.  In fact, unless a demigod's health isn't being altered at all by attacks, I don't see a situation in which that damage callback would make any difference at all?

What I'm getting at is that there has to be a better way to use that damage callback, say, to actually estimate incoming damage rates vs remaining health, but I don't know enough about the sensor/goal system right now to make that happen.  I think the health sensor is doing this same job somewhat ok, but it can't make projections w/r/t when it should retreat, it just reacts.

Quick test on v25n.

Standard test settings as per previous posts.

UB (human) / Sedna / QoT v UB (AI) / Sedna / QoT

The findings:

I wanted to run the test with two generals and one assassin on each team to see how it works. QoTs on both teams quite altruistically gave one monk to UB and one monk to Sedna, thus running with no monks at all, while Sednas were sporting an impressive retinue of 3 monks. After the monks died and were resummoned, the AI was happy just to leave monks attached to their original summoners.

Yeah, the damage callback is relatively unchanged from the original - if anything, it'll activate more often, but I still don't think it will be doing anything unless the main health sensor update happens a lot less than I'm assuming it does.

Well, yes, that's the only way to access table keys that aren't specifically named/numbered, but what's the context?

The shopping walking away and shopping again is caused by letting the AI decide what to do in the shop area.   Its either that or force them to shop and risk them sometimes not making it out of the shopping area.

I have not seen a consumable heavy purchase in a long time.  The priorities are such that every Nightmare AI should buy Nimoth first.  A healer idol or unbreakable boots and then some lesser armor items.  Bah just saw you were testing on normal.  I may need to redo priorities for a 1000 cash start.  I may have had tp above banded armor/helm.  Which would probably have it waste its cash on tp scolls and then pick up a potion cause it had nothing else it could afford.

Sound like a different game in your version.  Don't know if nightmare to normal would do all that.

There is some built in logic based on the targets health to continue an attack even if it is high risk.  That is probably the pursuit into enemy towers you saw.

If the DG takes damage it should drop its capture action and figure out what to do based on what is in front of it.  I will see if upping kill heroes helps and something to kill the weight of the capture action if enemies are nearby.  It may also just be too chicken to stick in a fight.  If you think it is that i can lower the bar for when it does a full retreat.

.23 or .24 added citadel response upgrades based on if the enemy gets certain troop upgrades, so its not giants vs soldiers.  They looked decent to me so havn't touched them.

Re: Citadel upgrades, Currency II is probably cheaper than Catapults.  Unless you set all priorities to zero (which it's basically doing except for FSI, creeps and CurI/II and ExpI/II), the AI will buy what it can afford.

Oh, I see the problem.  Someone threw out my changes to currency that took into account time, gold settings, etc.  With those settings, Cur II and Exp II were basically always 0 unless the AI was Hard/Nightmare, gold was high, or it was really early in the game (e.g. High War Score or something, where Cur II is available really early on and thus more worthwhile).

Edit: Replace the current block(s) of utility functions in AIGlobals with this: http://pastebin.com/yqFBKrNA

And then replace CitadelUpgradeWeights.CGoldIncome1-3 with this: http://pastebin.com/R8S6Dt1x

Uncomment the log message if you want to see what it's returning, when.

Note that there are also a bunch of other utility functions there.  I've had good results with allowing the AI to buy Graveyard I and II under extenuating enough circumstances, and I can post the priorityfunctions for those if you're interested.

Also, I get the feeling that you guys are testing this with different AI difficulties. 

We were testing with Normal AI, not Nightmare. However, the difficulty settings are consistent throughout both Pacov's and my tests (I based my settings on the tests we did together and his testing log).

I did like 3 games today with it (Solo, nightmare ai). They have improved, but it always seems like my ai are always doing dumber stuff then my enemy. The opposing ai will rush one lane together pushing me out, destroying a tower and taking a flag before my teammates seem to notice when im fighting a ai about to die who is sitting between me, and the tower andteammate Ai, The Ai will sometimes just stand there and attack creeps.

Nightmare has more gold than normal and can compensate.  My premise is that normal SHOULD BE ADEQUATE, and then nightmare should be challenging (and I'm quite willing to nerf nightmare until ws3 - then let them run wild - if the logic is good, the nightmare team should have more gold than the other and be able to buy better equipment and upgrades).  I've suggested logic that the normal AI focuses on critical upgrades in a timely manner, then buys cool stuff for itself.  As the nightmare AI has quite a bit more money at its disposal, its less of a problem and works with the existing rules better... but that's only because it has so much money.  Normal AI require the rules to be optimized to function appropriately.  At any rate, try a test with normal AI and let me know if I'm exaggerating.  That said, again, I want normal AI to cover the basic functions. 

If I really had to make a list of what I really want a terrible AI to do, it would be EXACTLY this and nothing else:

1. buy fs1
2. buy currency at ws3
3. spend all money at ws 8 on purchasing creeps

Everything - and I mean EVERYTHING else is gravy.  As it stands, item 1 is covered.  Item 2 is not.  Item 3 means that the ai will get priests then and little else. 

If I was optimizing for this base build, then I'd make sure objective 1 and 2 were always met.  Objective 3 requires additional logic (what is ok to buy if the ai is saving for those creep upgades). 

I'll start testing on normal first.  Probably got the priorities out of wack with that.  Just moving TP below banded helm and armor gets them buying the right stuff at the start.

Adding these conditions to the AI's item budget. 

Basically should be -1800 (cost of currency) once the AI reaches war score 2 and persists until the AI has currency.

It will start saving again when the enemy reaches 8 or it reaches 7.  It will save 4800 (cost of cats) until cats are purchased.

Final saving routine sets aside 6000 for giants at war rank 9 or above.

Now they AI at anytime can use all money on citadel upgrades, so if it ends up wasting money on lower upgrades a similar budget override may need to be applied to the citadel purchase side.   This should only affect the troop upgrades as the AI should only buy experience and FS1 before currency.

I'll test this a little bit and probably release it all as .26.

I would like to hear your criticism of the logic fleshed out a bit.

My take is gimping is cool if its done after the initial buy and prior to ws3.  Is that what you mean?  In a good game in MP, no one is dying before ws3, and currency is bought EXTREMELY close to ws3 coming up.  Curious if that is what you mean or if you are talking about somethign else.

If it becomes an issue we can probably modify the gold saving section i have now to check in on each ally brain's gold.  The ally brain with the highest gold can have its item budget reduced to purchase upgrades.   May be less reliable as the AI can still purchase other upgrades while its item budget is reduced.