Skirmish AI: E323AI 3.22.4

[Error323]

Heh cool, I know for a fact it can beat two KAIKS already. Just increase the minimal amount of scouts for each state within the config file and their economy will be raped hard.

[JohannesH]

Now, onto AI again. I noticed it has two basic flaws under BA 7.04. First flaw is connected with defense - what it chooses to build and when. Late game it basically doesn't build any defenses, while early game it builds those powerful anti-air towers which are simply unnecessary (few defenders/pulverizers and one flak in centre of base is more than enough to defend from early air rush of all kinds of air forces).

Second, more important flaw in my opinion, is that it doesn't play Core competitively at tier2 bots. It doesn't build Pyros which are great, genius, magnificent (lol!) fast assault, rush units. One unit of it can burn half of entire base in matter of seconds. It doesn't build Light Mortar units either, pretty much the only fast and mobile counter to spam of t1 bots and more powerful t2 arm bots. Sumo is simply a no-go versus Fatboys. Dominators alone are too inaccurate to deal with that.

What I propose is that you introduce vehicle plant after minor tweaks to defense, because Core tanks are much more mobile than their heavier kbots, what makes them perfect, all-round solution to deal with Arm. Few Reapers and Goliath should easily dispatch any Fatboy even in difficult terrain, Sumo, the Can and the rest not so much. Funnily enough, more of a problem of functionality than of balance in this case (you can't set a fair match in kbot vs kbot situation, as it's known core Kbots are inferior in skirmish).

Sorry but i dont think you know much of what youre talking about.

[slogic]

Bugreport. AI took 5 builders & move them across the map to build an LLT at (i guess) (0, 0) position.

[Error323]

E323AI v3.3.0 - High Templar
Changelog:

• Vehicle support
• More metal makers, less energy storages
• Less defense, no more AA defense
• Higher lateral dispersion for T3 units in groups
• Config update
• SIGSEGV fix
• No watermaps support yet

E323AI SO/DLL: E323AI-HighTemplar
E323AI SRC: E323AI@github

Note that the linux *.so file is for 64 bit systems only. This is why jeykey had problems.

[Achilla]

Now, onto AI again. I noticed it has two basic flaws under BA 7.04. First flaw is connected with defense - what it chooses to build and when. Late game it basically doesn't build any defenses, while early game it builds those powerful anti-air towers which are simply unnecessary (few defenders/pulverizers and one flak in centre of base is more than enough to defend from early air rush of all kinds of air forces).

Second, more important flaw in my opinion, is that it doesn't play Core competitively at tier2 bots. It doesn't build Pyros which are great, genius, magnificent (lol!) fast assault, rush units. One unit of it can burn half of entire base in matter of seconds. It doesn't build Light Mortar units either, pretty much the only fast and mobile counter to spam of t1 bots and more powerful t2 arm bots. Sumo is simply a no-go versus Fatboys. Dominators alone are too inaccurate to deal with that.

What I propose is that you introduce vehicle plant after minor tweaks to defense, because Core tanks are much more mobile than their heavier kbots, what makes them perfect, all-round solution to deal with Arm. Few Reapers and Goliath should easily dispatch any Fatboy even in difficult terrain, Sumo, the Can and the rest not so much. Funnily enough, more of a problem of functionality than of balance in this case (you can't set a fair match in kbot vs kbot situation, as it's known core Kbots are inferior in skirmish).

Sorry but i dont think you know much of what youre talking about.

Sorry but I'm afraid that you apply Player vs Player logic in an AI vs AI match. Oranges are not Apples.

[slogic]

After 3.3.0 release i started to test current AI vs 3 other allianced AIs in different combinations (2xKAIK+RAI, 3xKAIK). No, it does not win yet, but stands for about 25 min on Mount Dustmore which is reach enough with metal resources.

May i suggest building T2 lab in parallel with two defenses around (or move there any groups of units to guard building site) it because AI always decides to build it too far away from native base, so there is a 50/50 chance it will lost it later without getting any profit?

Why can't AI handle two or more labs of the same type? You may consider second fab as a backup fab :)

Often scouts behave strange when selecting next target. They are still in the safe place & can continue crushing everything around but they select another target far-far away instead.

[FluXy]

Can you make sure you installed the configs correctly and try other maps? Is there a case where it doesn't crash? (It really shouldn't crash )

It crashed around 40 minutes ingame.

I tried 3.3.0 now and it didn't crash this time but it lost

[Umrug]

Tried 3.3.0, BA 7.04, usual land maps

First - no signs of any performance issue at all. Played long game on big maps, with 100+ units - all went smooth and didn't crash. This outweighs everything I say below :)

For the record - the AI really plays worse now then it did (and I'll try to explain why later)

Some issues:
- minor - commander is idle for about 9 seconds on the game start (everyone needs to smoke a cig I guess)
- bug - (as someone said above) scout target selection - yeah, it does pretty stupid stuff, running from empty enemy base to some mex far away. Can you give it some priority for close targets? It worked OK once...
- Too many anti-air units in attack groups. There is really no need for 2 crashers(aa bot) in 5-bot group. I'm thinking about some ratio system here - can cost of all aa units in attack group not exceed some ratio (lets say 1/5, or per-mod defined, or, perfectly, dynamic value which grows if enemy has air units/labs?)
- No static aa at all - that's also bad. Same thing - can cost of all aa def not exceed some ratio (separate from the one above, just in case)
- Again, too little static defense. Is it build only when there is resource abundance? It should not...


- And now the real problem, for which I currently can think of no good solution. The problem - lets say at current state it has minScouts=6 and minWorkers=8. And at some point AI looses its entire attack group - meaning it has only scouts and static def left. Now, enemy still has mobile force operating, and its closing on your base. Scouts run towards enemy base, and die rapidly - either from llts or from the attack group. Same with cons - enemy group kills them. The problem is - AI keeps spamming scouts and cons only! It can not recover from that loss! Even when enemy group starts killing off resource buildings, and AI goes down a state - still, it will normally just produce stream of scouts that die uselessly. (Again, atm i don't see good and simple solution to this, I'll think)

New config (a bit more sane) for BA 7.04 - simply less scouts and cons, playtested, plays better.
http://rapidshare.com/files/304340134/B ... g.cfg.html

[yuritch]

AAI had the 'scouts/cons instead of combat units' problem in the past. It was fixed by introducing an 'emergency mode' (of sorts) - if the AI has too few combat units and the enemy has many, then scout/con production is halted (even if all of them are lost) and only combat units are built until situation improves. Something similar can be used here perhaps?

(this can backfire of course - if the enemy army turned back before reaching the base, AI will have a hard time rebuilding as it needs to get out of 'emergency mode' before being able to make cons to rebuild lost mexes. That can be solved if the AI knows the location of enemy forces to see that they aren't nearby anymore).

[Error323]

May i suggest building T2 lab in parallel with two defenses around (or move there any groups of units to guard building site) it because AI always decides to build it too far away from native base, so there is a 50/50 chance it will lost it later without getting any profit?

I'll try to fix this, but the algorithm I designed for it works a bit different and I don't wanna perform these kind of hacks.

Why can't AI handle two or more labs of the same type? You may consider second fab as a backup fab :)

Well, is it efficient really? It costs extra while it already builds 2 level 1 labs (veh and kbot). It would be better to invest in nanotowers imo.

Often scouts behave strange when selecting next target. They are still in the safe place & can continue crushing everything around but they select another target far-far away instead.

This is really weird, I didn't touch the scouting algorithm at all, what I did notice, is that fleas in BA are really bad their firepower and health is way to low. This is to everyone, if you change the categorization such that you remove "SCOUTER" from fleas performance increases dramatically.

[Error323]

Tried 3.3.0, BA 7.04, usual land maps

First - no signs of any performance issue at all. Played long game on big maps, with 100+ units - all went smooth and didn't crash. This outweighs everything I say below :)

Good!!!

Some issues:
- minor - commander is idle for about 9 seconds on the game start (everyone needs to smoke a cig I guess)

I've heard more ppl about this and will fix it.

- bug - (as someone said above) scout target selection - yeah, it does pretty stupid stuff, running from empty enemy base to some mex far away. Can you give it some priority for close targets? It worked OK once...

See reply above

- Too many anti-air units in attack groups. There is really no need for 2 crashers(aa bot) in 5-bot group. I'm thinking about some ratio system here - can cost of all aa units in attack group not exceed some ratio (lets say 1/5, or per-mod defined, or, perfectly, dynamic value which grows if enemy has air units/labs?)

I know what causes this and it can/will be fixed easily

- Again, too little static defense. Is it build only when there is resource abundance? It should not...

mmm yeah, its hard to find a proper balance between defense and offense. I don't have a real solution yet for this problem (open to suggestions).

- And now the real problem, for which I currently can think of no good solution. The problem - lets say at current state it has minScouts=6 and minWorkers=8. And at some point AI looses its entire attack group - meaning it has only scouts and static def left. Now, enemy still has mobile force operating, and its closing on your base. Scouts run towards enemy base, and die rapidly - either from llts or from the attack group. Same with cons - enemy group kills them. The problem is - AI keeps spamming scouts and cons only! It can not recover from that loss! Even when enemy group starts killing off resource buildings, and AI goes down a state - still, it will normally just produce stream of scouts that die uselessly. (Again, atm i don't see good and simple solution to this, I'll think)

The solution to this problem is definitly not trivial, and indeed difficult.
I'm not really sure if there is a proper solution at all.

[Error323]

Here is a config which wins of KAIK on atleast the following maps:

• Small Divide
• Barren
• Comet Catcher Redux
• DeltaSiegeDry

Check the difference and try it out if you want to.

[yuritch]

Well, is it efficient really? It costs extra while it already builds 2 level 1 labs (veh and kbot). It would be better to invest in nanotowers imo.

Making more labs is better if the units are built very fast. Take fleas for ex. - with a few assisters the lab spends more time waiting for completed fleas to leave it (so new ones can be built) than it spends actually building said fleas. Two labs in that case are an improvement. Probably same applies to pewees/flashes/etc. cheap fast to make units.
Another reason is security - two or more labs are harder to lose to a single bomber raid or some of that nature.

Also, when giving an assist order, be sure to check that target unit canBeAssisted is true. Not important for BA/XTA, but some other games have non-assistable factories, in which case assisting cons just surround the factory and do nothing :)

[slogic]

yuritch +1

Two (or more) factories of the same type allows AI to replenish your assault forces faster. You may consider it as a first step for solving problem Umrug mentioned about restoring AI's war spirit after successful enemy assault. Of course you should not build them immediately. Second (third) factory is needed later. This may depend on map size & timer.

Is it CPU intensive count enemy units?

About defenses. Make it appearance probability in buildqueues configurable & we decide what is the best during testing.

By the way, i like Fleas, they are fast & very irritating (for enemy). Just fix target selection for scouts :)

[hoijui]

about the second lvl2 lab:
i agree with Error323, it makes really no sense to build a second one as backup in a mod like BA. what yuritch explained is certainly right, but it is a totally different scenario (i am sure he knows that). The right thing to do is to make sure the lvl2 lab is better protected, eg. built in the main base or an other spot that is especially safe, or it has some static defence around or nearby units ready to defend it. for additional build power, nano towers should be used.

[slogic]

Yep, agreed. I was talking about level1 labs only.

[Error323]

E323AI v3.4.1 - High Templar
Changelog:

• Much better target selection
• Massive cpu load optimizations in large games
• Config update
• No watermaps support yet

E323AI SO/DLL: E323AI-HighTemplar
E323AI SRC: E323AI@github

[JohannesH]

Why can't AI handle two or more labs of the same type? You may consider second fab as a backup fab :)

Well, is it efficient really? It costs extra while it already builds 2 level 1 labs (veh and kbot). It would be better to invest in nanotowers imo.

Its efficient. Depends on what youre building (also a bit whether youre more short on energy or metal), you should usually build 2nd fac when you have 3-4 nanos usually. When you want to just keep building units from that fac type that is ofc, if you plan on adding another type of factory you should keep on adding nanos until you want to do that. And line the factories up so that nanos dont need to turn while switching between which fac to assist.

[Error323]

E323AI v3.8.2 - High Templar
Changelog:

• Wreckage reclaiming
• Pathfinder improvements (no more wait commands)
• Defense fix
• Better enemy micro
• Metal Maker usage improvement
• No watermaps support yet

E323AI SO/DLL: E323AI-HighTemplar
E323AI SRC: E323AI@github

If you guys don't find any weird crashes, Hoijui can add it to the installer of spring! :D

[A_Penguin]

I tested out the latest version on BA7.04, and came up with two problems:

-Metal maps appear to make the game crash at start (metal heck, speedmetal, speed ball)

-I played a map with no metal, and the AI made mexes until it boxed itself into the corner... it didn't make any metal makers.

It's a great AI. I beat it on Red Comet, but it didn't have too much trouble with me on Small Divide... I can't wait till it can do water maps :)