Skirmish AI: E323AI 3.22.4

[romulous]

3.20.0RC2 plays well. It's start game is good but the mid to end game still needs work. Would be good to see the bot tech up at some point and forget L1 units.

[edit] Ok on mount dustmore it played much better, sending T2 stuff [/edit]

[1v0ry_k1ng]

getting rl friends who play zero hour to try out spring VS this AI lead to some hilarious defeats. The mex raiding really makes people rage out

[Sgt Doom]

For some bizarre reason, some of the bots of this AI remain completely idle in XTA e.g. if I have 5 of them, only 3 will do something. Has happened to me on XTA 9.6, on 9.602 none of them do anything. Using latest one.
Using latest stable version of spring, this happens regardless of me using the multithreaded exe or not.

[Error323]

Hey Sgt Doom,

In both cases you should really check the logfile of E323AI and the spring logfile for clues as to why. The reason that the AI is idle in 9.602 is because it cannot find any configfiles for that version of XTA, you probably just want to copy them over from XTA 9.6 to 9.602 and it'll be good to go.

-- Error323

[slogic]

I know why AI can stall. This is because best scored target is not reachable by group, while AI does not check attack task was successfully created. And following iterations start with the same group and may last forever until target is dead or changed its position. I fixed this issue by random shuffling group IDs in a list to be scanned. Fix in my local master. But there is much work to do before uploading. So, do not expect too soon.

Here is an example why this happens:


AI's pathfinder takes start from group centre, not from position of one of the units in a group. So, it fails right at the beginning. Also unit may be pushed to a place which AI pathfinder marked as unreachable due to much less resolution of its graph.

[Sgt Doom]

Hey Sgt Doom,

In both cases you should really check the logfile of E323AI and the spring logfile for clues as to why. The reason that the AI is idle in 9.602 is because it cannot find any configfiles for that version of XTA, you probably just want to copy them over from XTA 9.6 to 9.602 and it'll be good to go.

-- Error323

Played with 3 AIs on Battlecity V1, XTA 9.6, both of the opposing AIs failed to do anything, the log is spammed with this:

[00:02] (EE): CPathfinder::getClosestNode failed to lock node(3,5) for pos(384,608)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(3,3) for pos(384,384)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)

Moved me any my ally to the other side of the map, same effect; my ally was unaffected but the opposing team was comatose.

Only seems to do this on XTA, which is a shame

Does the same on XTA 9.602 with the configs for 9.6 copied over.

EDIT: If I enable the Xtaids unit pack and/or disable the XTA expansion pack, my ally does not do anything and the other team becomes active.

[Zydox]

I've been playing a few "3 E323 AI against me" battles... but the last three all crashed... here's the latest crashreport... AI logs, infolog & demo... http://www.zydox.se/Crash20100510.zip

[SirMaverick]

[00:02] (EE): CPathfinder::getClosestNode failed to lock node(3,5) for pos(384,608)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(3,3) for pos(384,384)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)
[00:02] (EE): CPathfinder::getClosestNode failed to lock node(4,4) for pos(512,512)

Moved me any my ally to the other side of the map, same effect; my ally was unaffected but the opposing team was comatose.

Only seems to do this on XTA, which is a shame

I've seen this in BA and CA, too.

http://springrts.com/phpbb/viewtopic.ph ... 94#p426894

[1v0ry_k1ng]

urgh how can you use platters

SO ugly

[Jazcash]

urgh how can you use platters

SO ugly

I use platters with near to the lowest graphics settings possible
If I wanted to play for graphics, I wouldn't play Spring at all.

[AF]

Platters are a usability/visibility thing, not an aesthetics thing. I use them because it means its easier/faster to tell apart different teams at a greater distance. I can pull out and see what units belong to who and most importantly, I can visually identify each unit type by sight, rather than indiscriminate radar blips.

[slogic]

New version 3.21.0 RC1.

Changelog:

• refactored target selection
• fixed most issues with attack stalling
• avoid continuous issuing of MOVE command in Merge task when group is already in-place
• changed Merge task algorithm: groups try to merge in safe places and closer to enemy at the same time
• changed factory building strategy to tech-up faster
• detection and resolving of factory stall
• WIND & TIDAL tags removed
• fixed multiple AI info printing in chat
• added support for chat command "!e323ai" to visualize debug info (available for spectators only)
• added support of separate categorization files per AI; you have to copy & rename original categorization file to the following template: <modname>-categorization-<team_id>.cfg

DLL is compiled from "Debug optimized" build, so on crash infolog.txt should contain stacktrace (without source line numbers though) with full qualified function names. The only requirement is keeping SkirmishAI.pdb file near spring.exe.

Please, test and report crashes if exist.

[Zydox]

M'kay... here's my first crash with the new AI version...

[MisterBenn]

Well done guys I think this new release is a big step forwards! I've played 3 BA games and a 45 minute 3v3 with SA so far:

• No stability issues for me so far.
• Whatever pathfinding and/or target assessment issue that was making attack groups coalesce and dither on the battlefield seems to have been fixed! The AI plays a decent game of SA once again.
• Teching up is more reliable now for sure. The aircraft addition had made teching slower as a third T1 factory seemed to take precedence before. Haven't seen an aircraft in my games yet with this version but the progression to T2 comes more quickly and reliably now it seems.
• Turret building is looking quite good too compared to a few versions ago! Don't know when this improved (or whether its my imagination) but on the Folsom Dam 3v3 I played, both sides had clustered turrets around the sloped base entrances to good effect. In response to where attacks occur? Perhaps some heavier turrets would occasionally be good? Pit Bulls / Vipers / Popups etc go a long way, the AI has a strong preference for LLTs throughout the game unless it becomes super rich. Curiously on my BA FFA on The Cold Place, all 3 sides made a beamer as the first turret!

Some suggestions:

• Should Sniper and Artillery tagged units be more protected from damage? Sometimes they hang back and fulfil their role well, sometimes a sharpshooter is assigned an attack order on a single windgen on the opposite side of the base and tried to walk straight through.
• Do you have some nano logic on your list? Even some simple orders would go a long way I think, like giving a patrol order and then making them wait if stalling is occurring. Some nano factory assisters would be great of course and might help with late game base congestion.
• As was mentioned before, factory positioning is a problem sometimes. In 1v1 on Alien Desert, the player starting in the bottom left built his factory up against that monolith three games out of three and couldn't get his units out.

Please keep it up though, this is my favourite AI :)

[slogic]

M'kay... here's my first crash with the new AI version...

I need the log file. I think some unit became idle which was not registered by AI. Anyway i know how to fix this. Just curiosity.

PS. By the way, your game is full of desyncs. Good stress for AI and Spring engine itself

[slogic]

Teching up is more reliable now for sure. The aircraft addition had made teching slower as a third T1 factory seemed to take precedence before. Haven't seen an aircraft in my games yet with this version but the progression to T2 comes more quickly and reliably now it seems.

Before 3.21.0 factory build list was: kbot#1, vehicle#1, aircraft#1, kbot#2, etc. Now it is kbot#1, kbot#2, kbot#3, vehicle#1, vehicle#2, etc. If after kbot#3 vehicle#3 (or #2) is directly available, it will be built. Also don't forget there is a tech limit in config file per resources income. If you think that after kbot (or vehicle, depends on what is calculated as primary factory) should follow aircraft, then vehicle, then elaborate on this. For me it is not evident.

Turret building is looking quite good too compared to a few versions ago! Don't know when this improved (or whether its my imagination) but on the Folsom Dam 3v3 I played, both sides had clustered turrets around the sloped base entrances to good effect. In response to where attacks occur? Perhaps some heavier turrets would occasionally be good? Pit Bulls / Vipers / Popups etc go a long way, the AI has a strong preference for LLTs throughout the game unless it becomes super rich. Curiously on my BA FFA on The Cold Place, all 3 sides made a beamer as the first turret!

More defence buildings appeared since 3.20.0 with a little tweak in defense matrix module. What types to build is almost a mystery to me cause i did not dig it. It is related to algo which calculates the affordable state of unit being built. Defence upgrading will not appear until tasks queues are implemented (see issue tracker).

Should Sniper and Artillery tagged units be more protected from damage? Sometimes they hang back and fulfil their role well, sometimes a sharpshooter is assigned an attack order on a single windgen on the opposite side of the base and tried to walk straight through.

I understand this. Long ranged units should not be mixed with short ranged ones because in the end of their task long ranged start to shoot while other units in disorganized order still moving to their shorted range. Also they should have different behaviours on battlefield. Current AI design has 3 tactics: harass (scout) engage & expand. Engaging is done via groups of MIXED unit types, so attack algo is COMMON for all of them. Here is all that AI can boast.

Do you have some nano logic on your list? Even some simple orders would go a long way I think, like giving a patrol order and then making them wait if stalling is occurring. Some nano factory assisters would be great of course and might help with late game base congestion.

Error should deal with it. When? I don't know. Tease him, beg him or think up anything else. By the way, this problem is on the issue tracker list (#2). So, please, don't post everything which is in the issue list already. That's boring

As was mentioned before, factory positioning is a problem sometimes. In 1v1 on Alien Desert, the player starting in the bottom left built his factory up against that monolith three games out of three and couldn't get his units out.

Issue#36.

[MisterBenn]

Hi again!

...then elaborate on this

Sorry for not being clear, I was just trying to say that in the previous build, the fact that AIs would build aircraft plants as a third factory before teching up was slowing down their pursuit of T2 more than before. That's in the past now, I ran some more games and the teching up is vastly improved.

I'm also liking a lot about the attack algorithms now. I like how the battle front retreats when they are taking damage, you get a shifting battle front and often when they retreat they end up merging with a new group joining the front and then push back again! It's a nice touch. Also I saw some effective bombing runs taking place.

I'll take a look at the issue list again before posting issues, originally I wasn't sure whether issues were best posted here, whether you wanted the issue list for developer issues rather than user issues etc. I'll take a look there from now on.

BTW, I saw that Supreme Annihilation U42 was released, I attached a config file again. It's very (very) similar to the BA config files now - you might like to make the default configs that E323AI generates have the tags currently used by your BA config as they are better in all cases I think.

Anyhow, please keep up the good work, the more I play this the more I like it!

[jseah]

Very nice on the new version. It's definitely more stable for me.

Is it possible to have some kind of function to turn off the cheating part of the raiding? Mainly because playing 1v1 on large maps meant for 4v4, I find I have to attack early to win at all. Early bombers or gunships mean a boring 5min game (I just bomb the factory to stop the AI building anything but scouts)
Sometimes I get away with porcing in my base and having a mini army defend every extension. Doesn't work all the time, strongly depends on getting enough mmakers.
Another mid game winning strategy is building an aircraft plant and peepers on repeat to the back of my base to bait the AI into wasting tons of resources in building AA when it hasn't seen a plane yet... (hey, if it cheats, so can I... right? XD I note that it never builds AA until someone gets an aircraft plant)

The micro management for defending against raids once you get more than 10 mexes becomes impossible without turning down the speed to 0.3, and it only takes so much micro because the AI's A.K.s can magically see where my samsons are. (it's still possible to catch them due to superior range and flanking, but it takes ~20 commands per A.K.)


The entire scouting/raiding thing fails when playing on certain maps. Maps that have three or less channels for land units, say. DSD is one. (1v1, I can win easily by dropping LLTs in the mountains and pushing bottom, 4v4 is similarly easy) Supreme Battlefield is another for obvious reasons.

In CA, the raiding with 1 bot doesn't work if I go gunships at the start. CA t1 gunships are too accurate to jink (arm's is a laser) and they move faster than raiders (obviously) Perhaps it could use T1 gunships to raid? XD

[slogic]

> It's definitely more stable for me.
I was not focusing on stability though :) Just programming and testing.

> Is it possible to have some kind of function to turn off the cheating part of the raiding?
Turn off scouting? Yep, it will be possible when difficulty levels are implemented.

About Peepers usage for win strategy. Sounds as nonsense because they are not tagged as ATTACKER and so are not counted as markers for economy to build anti-air units. At the same time Peepers are treated as helpless targets and being behind your base they lead to issue described below. So, problem should be not of the quantity of Peepers, but placement of helpless units.

> The entire scouting/raiding thing fails when playing on certain maps.
Yep, i know. AI needs to control threat on all path segments to the target. Or need some threat threshold value to prevent generating such path if some segment is above this threshold. Next step is to mark this area for ENGAGE groups to clean up a path for SCOUT groups. Etc., etc., etc.

> In CA...
At current stage I personally don't care about CA, because it uses plain tech tree which needs another approach in AI planning. Also no config files for CA are deployed with AI.

[jseah]

>Stability: Well, it stops crashing when the commander dies... So that's more stable. Haven't tried multiplayer to test whether it desyncs as often.

>Peepers: Could it be registering my con aircraft? I only do a bunch of bombers near mid to late game. Although the bombers do outnumber my army + defense at least 3 to 1.

I do notice that AA only appears when I build air, and never otherwise, even if the AI nevers see a single air unit. Just thought I would exploit that behaviour to get the AI to waste m building AA. It did work to a certain extent in that attack groups became about 1/4 AA which took some pressure of my ground forces.

------------

Just played a game taking my time in supreme battlefield. Nice to see some air in there. =) Any way to stop T1 bombers running into a fighter screen?