Enhanced AI Mod v0.26.35 Release 2/16/2011

skirmish (mod version) .26j/uberfix 1.03 (last version published by miriyaka)

Settings: cata/High Towers - all other settings on default
AI settings: nightmare
2v2

UB(h)/sed v da/reg

The purpose was to test the three DGs whose skills use was altered significantly in the last version.

The findings are:

(1) Sedna's Silence use is still very bad. She should never use it at the very start of the 1v1 combat while both are at full hp. Period.

(2) DA's Swaps are MUCH better and much less random-looking. Good work.

(3) Reg's Snipe became quite dangerous in this version (there is plenty of those moments demonstrated in the replay). No more point-blank Snipe nonsense, too. Very nice, indeed.

(4) Overall AI logic is good. Ports to save allies (me) and priest sharing were pretty much decent. The AI even fake-ported to the tower we were destroying at certain point (at least that's what it looked like!).

(5) Cit upgrades. Once again, the enemy AI didn't get ANY creeps. Late game, it bought Bracelet of Rage for the Reg, but no creeps. Not even priests. My AI was doing great, though: the basic three upgrades + all the creeps. The replay is here: http://www.gamereplays.org/demigod/replays.php?game=51&show=details&id=171064

Successfully updated karma reason!

TY

Is it possible to make some demigods primary savers. Maybe allow them to spend a maximum amount of gold (say 6000) on items before they begin to save (this would allow reg/tb with BoTS to have infinite mana and about 5000 health - Unbreakables, Nimoth, Banded, Scaled + 1 Helm, several teleports and potions.

IMO, Reg and TB ;+ blade of the serpent build;(maybe QOT with shield/ground spikes build and lots of mana items) should nearly always be primary citadel purchasing demigods, as creeps reward these low life/DPS characters the most and have great AOE abilities which can be used regularly to farm gold. Also the bonus gold that the AI gets would allow these demigods to reach these thresholds quickly.

This would also allow demigods which do better when upgraded with artifacts; UB/LE/Rook; to purchase these items more often (provided there was already a saver on the team).

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

(3) Reg's Snipe became quite dangerous in this version (there is plenty of those moments demonstrated in the replay). No more point-blank Snipe nonsense, too. Very nice, indeed.

'

exciting news. Good stuff!

(4) Overall AI logic is good. Ports to save allies (me) and priest sharing were pretty much decent. The AI even fake-ported to the tower we were destroying at certain point (at least that's what it looked like!).

also good!

Successfully updated karma reason!

Quoting IN1,

reply 126

skirmish (mod version) .26j/uberfix 1.03 (last version published by miriyaka)

Settings: cata/High Towers - all other settings on default
AI settings: nightmare
2v2

UB(h)/sed v da/reg

The purpose was to test the three DGs whose skills use was altered significantly in the last version.

The findings are:

(1) Sedna's Silence use is still very bad. She should never use it at the very start of the 1v1 combat while both are at full hp. Period.

(2) DA's Swaps are MUCH better and much less random-looking. Good work.

(3) Reg's Snipe became quite dangerous in this version (there is plenty of those moments demonstrated in the replay). No more point-blank Snipe nonsense, too. Very nice, indeed.

(4) Overall AI logic is good. Ports to save allies (me) and priest sharing were pretty much decent. The AI even fake-ported to the tower we were destroying at certain point (at least that's what it looked like!).

(5) Cit upgrades. Once again, the enemy AI didn't get ANY creeps. Late game, it bought Bracelet of Rage for the Reg, but no creeps. Not even priests. My AI was doing great, though: the basic three upgrades + all the creeps. The replay is here: http://www.gamereplays.org/demigod/replays.php?game=51&show=details&id=171064

Nice game. Reg almost had you a few times.

Sedna silence, think i used the wrong function to check number of enemies. I'll try the counting blips function rather than threat level.

4) fakes teleport is more likely A.D.D than anything else. Probably higher priority target came by and interrupted the action.

Think the enemy team upgrades didn't go in response to yours because the save script doesn't release funds until war rank 8. I'll tweak it to allow the AI to spend its saved money in response to enemy upgrades.

Thanks for the report and the replay.

Successfully updated karma reason!

Couple comments:

Tested a 3 human vs 5 nightmare ai. It was decently challenging in the beginning, but after we passed a certain hump (around lvl 10) the game became dramatically easier, with kill after kill after kill (like 50 kills for our team after this point). Prior to, we actually had 2 deaths for our team. I don't specifically know what would make them better mid-late game, but they seem to be lacking, as the difficulty changed dramatically.

Also, I don't think reg builds should have bots unless they go at least one rank of wings. Mines don't work with bots so.. it really doesn't provide much of a benefit for him at all outside of wings. Staff of renewal, swift anklet, or wrath (only if it is intelligently used to kill creeps, help get assists / kills) would be much better choices.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

skirmish (mod version) .26j/uberfix 1.03 (last version published by miriyaka)

Settings: cata/High Towers - all other settings on default
AI settings: normal
3v3

rook(h)/reg/reg v reg/reg/reg

Observed results:

1st - both regs on my team went sniper/mines builds, so I should be seeing some sniping now. Very quickly saw 2 of the enemy regs focus snipes on one of mine. Good stuff.

saw 1 of my regs do a point blank snipe against another reg that had full hp.

The ai is still willing to run past the towers on the right or the left to go for the flag further down. It really needs to be adjusted so that the tower is the number 1 priority if its still standing on cata.

2 regs on the other team came pretty close to me and then sniped (pretty much point blank (but it wasn't a horrible move considering the damage they do early game and they stayed out of my tower farm... but still... point blank...)

ah - then the ai got me with a double snipe as I wasn't being cautious enough - good!

late game - ai appeared to have no interest in using mark of the betrayer... this skill should be used VERY often.

so... I got to thinking - maybe don't adjust the point blank snipe thing too much more now. I also still have the impression that the reg's aren't sniping enough. I don't think they really consider sniping as a huge priority unless someone is close. In a MP game, regs will often prioritize sniping WHOEVER is visible to them and just spam away on snipes. Anyway, just some quick thoughts.

Successfully updated karma reason!

Nice game. Reg almost had you a few times.

Yep. Through quite intelligent use of Snipe.

4) fakes teleport is more likely A.D.D than anything else. Probably higher priority target came by and interrupted the action.

I guessed that much But since that's what most people tend to do in MP when they leave tower unguarded, it was funny to observe

Thanks for the report and the replay.

My pleasure.

Successfully updated karma reason!

miriyaka,

trying to use the EnemyHasUpgrade and it always returns false. Do i have an old/wrong version in the AI globals file?

Found the issue, old:

Code: c++
function EnemyHasUpgrade(unit, upgradeName)
    local hasupgrade = false
    local etBrain = GetEnemyTeamBrain(unit)
    if etBrain then
        local hasupgrade = HasUpgradeByTeamBrain(etBrain, upgradeName)
    end
    return hasupgrade
end

hasupgrade was declared twice and the one in the if block didn't get returned only the one in the outer bock (always false).

Successfully updated karma reason!

Quoting Exxcentric,

reply 127

TY

Is it possible to make some demigods primary savers. Maybe allow them to spend a maximum amount of gold (say 6000) on items before they begin to save (this would allow reg/tb with BoTS to have infinite mana and about 5000 health - Unbreakables, Nimoth, Banded, Scaled + 1 Helm, several teleports and potions.

IMO, Reg and TB ;+ blade of the serpent build;(maybe QOT with shield/ground spikes build and lots of mana items) should nearly always be primary citadel purchasing demigods, as creeps reward these low life/DPS characters the most and have great AOE abilities which can be used regularly to farm gold. Also the bonus gold that the AI gets would allow these demigods to reach these thresholds quickly.

This would also allow demigods which do better when upgraded with artifacts; UB/LE/Rook; to purchase these items more often (provided there was already a saver on the team).

I thought about doing upgrades by build, but thought managing all the scenarios would be a headache.

Already have items by build, so can drop vlemish from builds with bots for example. I have that in on one of the UB builds. If that is true for all of them using bots i can drop it for them as well. Think plenor is still in though. Anyway, Can balance items by build as needed.

It may naturally happen the way you describe anyway, as every time it shops it checks if it is the highgold AI on the team and if it needs to save. So the builds better at farming should rise to the high gold slot.

Successfully updated karma reason!

Quoting 4Nana,

reply 130

Couple comments:

Tested a 3 human vs 5 nightmare ai. It was decently challenging in the beginning, but after we passed a certain hump (around lvl 10) the game became dramatically easier, with kill after kill after kill (like 50 kills for our team after this point). Prior to, we actually had 2 deaths for our team. I don't specifically know what would make them better mid-late game, but they seem to be lacking, as the difficulty changed dramatically.

Also, I don't think reg builds should have bots unless they go at least one rank of wings. Mines don't work with bots so.. it really doesn't provide much of a benefit for him at all outside of wings. Staff of renewal, swift anklet, or wrath (only if it is intelligently used to kill creeps, help get assists / kills) would be much better choices.

Up till warrank 8 the AI should play pretty safe. Mostly creeping and keeping flag balance in check. At 8 it gets more aggressive and tries to get a portal advantage, is this when you got the kills?

There are two reg builds:

- auto_attack - goes bots and angelic fury is maxed by 10. no mines or snipe in this build.

- Sniper_Mines - goes staff of renewal and maxes snipe and mines by 10.

Wrath breaks the AI both mine and the base game if it takes it, so it is out of each build. If it turns out to be critical to any build I can try to fix whatever is breaking it.

I can add multiple favor items to a build for it to randomly select from each game.

Successfully updated karma reason!

Also, I don't think reg builds should have bots unless they go at least one rank of wings. Mines don't work with bots so.. it really doesn't provide much of a benefit for him at all outside of wings. Staff of renewal, swift anklet, or wrath (only if it is intelligently used to kill creeps, help get assists / kills) would be much better choices.

I disagree, if you can put 1 point in wings for this build early on, BoTS pays of dividends for both mines and snipe (if you can keep reg with BoTS in wings form (which already seems to be happening). This is even more profound if you save mines for towers/demigods only --> farm with AA wings lvl 1

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

Also, I don't think reg builds should have bots unless they go at least one rank of wings.

probably best to just change it so AA build regs just stick with blood.

so... I got to thinking - maybe don't adjust the point blank snipe thing too much more now. I also still have the impression that the reg's aren't sniping enough. I don't think they really consider sniping as a huge priority unless someone is close. In a MP game, regs will often prioritize sniping WHOEVER is visible to them and just spam away on snipes. Anyway, just some quick thoughts.

The ai is still willing to run past the towers on the right or the left to go for the flag further down. It really needs to be adjusted so that the tower is the number 1 priority if its still standing on cata.

regarding that last point, I'll try a few more games with this version, but I think there needs to be some improvement. I'd go so far as to say if a reg has a snipe and the mana to cast, use that unless super close range (unless it would get a kill).

Successfully updated karma reason!

Quoting Exxcentric,

reply 136

I disagree, if you can put 1 point in wings for this build early on, BoTS pays of dividends for both mines and snipe (if you can keep reg with BoTS in wings form (which already seems to be happening). This is even more profound if you save mines for towers/demigods only --> farm with AA wings lvl 1

I said unless it gets one point in wings.. so Also, agree or disagree the fact remains that mines don't give mana back with bots.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

Please take a moment to watch this replay... it showcases some fundamental problems...

http://www.gamereplays.org/demigod/replays.php?game=51&show=details&id=171154

you might need (probably do) to have favor mod 2.0E installed to view this.

Don't bother watching anything early - just skip ahead to about 11 min 20 seconds. At this point in the game, I'm going out of my way to get snipe killed or even sniped. At about 11:30, the ai with snipe has me in line of snipe (and snipe range) but chooses not to take a shot even though at full mana and not on cd. At 12:04, I'm still dancing around, still in los for a reg with snipe (me at like 600 hp) and the ai won't even look my way. At 12:26, I'm taking their portal with 600 hp. 1 of their regs with snipe is at the crystal and has los, does not bother to shoot. 12:32, 2 regs have los with snipe and available mana and do not bother to snipe (I have 615 hp). 13:05, 2 regs have los with me at 700 hp and do not bother to snipe. From then until about 13:40, 2 regs have los on me and do not snipe (I'm around 1k now).

A snipe reg chases me for just a moment at 13:41 but decides he is better off attacking 1 single tower I dropped instead of pursuing me (I have 729 hp). He is also more interested in that than sniping - ai has tracker. The other reg that has los is not looking to snipe me now either.

14:26, I have 900 hp... a reg has los but chooses not to snipe me while i take their portal.

Anyway, just a sample... it seems like the ai refuses to snipe if its doing something else quite a bit or on its way somewhere.

edit - another thought - it might be good to start a few lists of things that need tweaked/bugs. A list for basic AI logic and a list for dg specific logic. I'll try to make a note in my replays of where requested tweaks, changes reside.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

skirmish (mod version) .26j/uberfix 1.03/favormod 2.0E

Settings: levi/High Towers - all other settings on default
AI settings: normal
5v5

reg(h)/4 regs v 5 regs

Observed results:

The Ai's actually coordinated pretty well for an early kill via snipe on levi. Generally speaking, the ai reg's seem to do better on a larger map... I'm thinking it may have something due to the distance to flags on cata

Basic AI Logic tweaks - no suggestions based on this game.

DG Specific Logic tweaks - check reg snipe logic on smaller maps using the replay I posted previously... I'd also recommend setting the priority for usage of mark of the betrayer so that it fires every single time a reg that has that ability is in range. I'd very much prefer to see that as a regs opening move if close enough to get into combat.

Successfully updated karma reason!

26k:

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26k.exe

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26k.zip

Fix buying troops in response to enemy, fixed function and made it a drop everything and go buy upgrades override.

Reg more snipe happy... Upped betrayer weight, not sure if it will do anything. Skill has no conditions, so should fire whenever its in range.

Sedna silence fixes/tweaks.

Difficulty on snipe i think is the AI is usually in different action or goal set when it could snipe. Most of the goal sets breakout when an enemy demigod is in range. That range is about a ranged attack distance x 2. Snipe is like attack distance times 10. to try and work with this i have made snipe and action valid for any master goal, and given it some weights in each goal category (higher weight actions get selected first/more often). I don't know how well this will work, but we will see.

+1

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

Thanks again, peppe - will give it a go.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

skirmish (mod version) .26k/uberfix 1.03 (last version published by miriyaka)

Settings: cata/High Towers - all other settings on default
AI settings: normal
4v4

reg(h)/reg/reg/reg v reg/reg/reg/reg

Observed results:

the ai reg still had problems deciding when to snipe. It missed any easy kill where 1 ai only had to chase 1 aa more (instead let the other ai leave safely).

I was attacking an ai at mid while THREE ai's were sitting on hp with snipes off cd and did not snipe.

noticed that if ai reg is on low hp (well.. below the threshold you set, ai will refuse to snipe even if a kill is available). You know how sometimes the ai will drop just behind a tower and then tp out? I'd like it if that logic was changed for reg so that 1)if a kill is available (eg enemy hp <400), then it will try to snipe regardless of its hp and 2) that the ai will throw up a snipe as soon as it reaches that safe distance behind a tower.

Another thing - I think the ai reg doesn't work like a human in another aspect. For instance, if I see a dg I want to snipe, even for a moment, I will lock my snipe on them. If they hit the fog of war, if I have already locked my snipe on, I will still get the snipe. I think that because the ai only considers snipes that are in range for it (THIS IS A FLAW WITH THE AI), then it won't do this. That's generally about 30-40% of the kills a reg gets per game lost right there.

Another thought - based on how I'm seeing the ai regs play, I suggest that all builds only purchase Blood of the fallen. They typically fall below their allotted damage threshold and run off, when they could do much more good if they stuck around.

so - cata still played piss poor with ai regs - they did SEEM to use mark a little more at least... but snipes were QUITE lacking

Successfully updated karma reason!

I think I found the final snipe issue. There is a 'nearbyheroes' variable that most skill reference to see if there is a target in the range of their skill. Well that variable caps at 30 yards. So when the snipe skill asks is there anything withing 90 yards of me, the variable often returns 0 cause it doesn't look between 30-90 yards.

I've got Reg sniping to near the center from his health stone now

The AI can shoot into the fog like player it just has to start the action before the unit is in the fog.

Or are you saying that you might be say 120 yards away and select your snipe target, so your unit now tries to close 30yards to shoot. Meanwhile the target moves into the fog. Your unit continues and snipes into the fog a unit now hidden?

I don't know if the AI can do that as it usually disables moving when its going to use a skill.

On the retreat vs stick around I can try to adjust all that. I got them to hang around the towers by removing run to health stone as an option until their health % is low. Giving them blood would just delay slightly when they reach that threshold.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

Or are you saying that you might be say 120 yards away and select your snipe target, so your unit now tries to close 30yards to shoot. Meanwhile the target moves into the fog. Your unit continues and snipes into the fog a unit now hidden?

Yes - exactly that. AI logic might be a bit tricky here, though. Some examples... let's say I'm on the mana flag on cata. I'm about to head over to the xp flag. Often, if I see any dg on hp, even for the briefest moment, I will lock a snipe onto them (while they are still out of range if I'm not close enough.

Another, much more practical example is this: Let's say an enemy dg is down to like 200 hp and is now running like hell to get to the crystal. In instances like this, I will immediately lock on and see if I can catch them with a snipe before they make it back to the crystal. This can mean walking through towers sometimes or whatever is necessary to get the kill shot off (preferably not dying myself, of course). This type of kill is how most good reg's get their kills in MP.

Successfully updated karma reason!

26L:

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26L.zip

http://peppe-dg-mods.googlecode.com/files/Skirmish_AI_v26l.exe

More Reg sniping tweaks.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

skirmish (mod version) .26l/uberfix 1.03

Settings: cata/High Towers - all other settings on default
AI settings: normal
4v4

reg(h)/reg/reg/reg v reg/reg/reg/reg

Observed results:

wow - MUCH better behavior. Saw a few instances where a low hp reg might have gotten away but the ai often went for the kill. It wasn't 100%, but SOOOOOO much better than previous versions.

Even saw one instance were a dg was about to escape to the crystal and my ai reg picked him off.

basic ai issues:

ai at one point just stood in front of his tower and just absorbed damage. Didn't try to flee, snipe, etc. Only happened once that I observed.
ai still wants to go for the gold and cooldown flags on cata while the tower is still there. Tower should be the first priority. It may have been slightly better in this version than in previous, but maybe not.

Overall, though - this was a MUCH better experience with the reg ai at least on cataract. I'm curious to see how the reg ai does on other maps now, Great job again, Peppe!

Successfully updated karma reason!

Testing Sedna Silence fixes/tweaks as announced in v26k + assorted issues.

skirmish (mod version) .26l/uberfix 1.03

Settings: cata/High Towers - all other settings on default
AI settings: nightmare
5v4

sed(h)/sed/sed/sed/sed v sed/sed/sed

Observed results:

This setup allowed the game to be long enough for our testing purposes. I sat at crystal as an observer and bought Cur2, priests, angels and giants. That's it.

(1) Each Sedna gave me one priest. Why is this? It will give us an asymmetrical picture in any game with human and more than 2 AI generals. Definitely needs to be tweaked: refine conditions for allied generals > 1.

(2) Some Sednas were prioritizing Minotaur minions over Demolishers. It should never happen. Priests > Demolishers > Minotaurs is the way to go.

(3) Immediately upon WR3 announcement, ALL of my AI Sednas simultaneously rushed to the cit in caravan-like formation to buy Cur1. I don't think it makes sense.

(4) My AI didn't get XP1, until I bought it at WR9 or so. No idea as to why did it happen.

(5) Finally, Silence fix feedback. During some prolonged Sedna skirmishes in the center and on the mana flank (roughly from 8:00 to 14:00), I've seen some bad (old) uses of Silence as a nuke. Also, seen one of the Sednas apparently doing something I'm so adept at: hitting 4 instead of 3, so she just spent mana on Silence when no one was even remotely in range No kidding! Check it out in the replay.

However, and that's definitely a great success, ~50% of Silences were timely. I've seen it used both to secure kills and as an escape option. Good job!

Replay: http://www.gamereplays.org/demigod/replays.php?game=51&show=details&id=171396

PS: What are the existing Sedna builds and buying priorities late game (especially artifacts)? All Father Ring I can surely understand, but Bracelet of Rage? Why?

PPS: Pounce is also used sub-optimally. IMHO, should give higher priorities to its use as an interrupt only, especially for Silence/Heal build, that doesn't level up Pounce until very late game. Hitting for 400 damage and spending some mana (even though it's dirt cheap) on it is idiotic, when you can hit for more damage with a regular AA + Nature's Reckoning.

Successfully updated karma reason!

pacov

Join Date 02/2008

+183

skirmish (mod version) .26l/uberfix 1.03

Settings: cata/High Towers - all other settings on default
AI settings: normal
1v3

rook(h) v reg/reg/reg

Observed results:

Overall it was fun. The ai kept up the snipe pressure for the most part forcing me to put more than a few points in structural transfer to keep up.

basic ai issues:

I died once late game (me ws10 them ws 8). I had taken their valor flag about a minute prior to my death. The ai did not attempt to retake valor. As far as late game priorities go, holding the valor flag is only slightly LESS important than keeping your own 2 portals. The ai had at least 35 seconds to retake the flag and still has not
At ws8, i did not see any of the 3 ai's porting in to keep me from taking their portal. Wondering if their might be a slight conflict with the snipe logic over tping in

Successfully updated karma reason!

Another thought: I think the modders, with all due respect, do not firmly understand how should Sedna's skills (Silence and Pounce) work in MP. It's not meant as a critique (kind of goes without saying it is not ), rather as something that seriously hurts the progress in that direction. My proposal is: maybe it will be beneficial to post a few replays of really good Sednas in various MP settings (2v2, 3v3) for Peppe? I have one or two replays (from 1.3 era) involving hedgie, and those Pacov v McOrcenz tourney games are also very representative. I have a few half-decent replays myself, but I don't want AI Sednas learning my favorite trick (wasting mana on Silence when no one is in 150 yards radius at least ), so I think I'll pass.

Successfully updated karma reason!