Skirmish AI: E323AI 3.22.4

[jseah]

It works now. I went through the config file and the logs and added tags from the wishlist errors (ASSAULT and SNIPER) using my own estimates.

The AI attacked across the sea and river using hovercrafts! Caught me by surprise there and nearly finished me off. Had to scramble gunships to counter.

That said, the raiders it uses get flattened by gunships.

Problems:
1 minute stall at the start is poor.

Code: Select all

[00:24] (II): ATask::remove BuildTask(9) EPROVIDER(Solar Collector) ETA(129) timer(130) Group(0): amount(1) [Commander(6459, -1)] Assisters: amount(0) [] 
[00:24] (II): CTaskHandler::remove BuildTask(9) EPROVIDER(Solar Collector) ETA(129) timer(130) Group(0): amount(1) [Commander(6459, -1)] Assisters: amount(0) [] 
[01:32] (II): BuildTask(16) FACTORY(Infantry Bot Factory) ETA(129) timer(0) Group(0): amount(1) [Commander(6459, -1)] Assisters: amount(0) [] 

I wonder what's it doing in between the 24s and 1min32s.

[slogic]

May be looking for vehicle or kbot factory depending on map height variance. Factory selection is too hardcoded currently.

[jseah]

slogic: Any way to solve this? At the current point, it's just about a match for RAI, when it should be completely owning it from playstyle.

General:
Played 1v1 BA7.11, DSD, me ARM vs AI CORE.
- Scouts too easily intercepted. AKs also fail vs LLT in the mountains.
- - Have scouts retreat if there's no way in instead of trying to rush an LLT alone

- Static samson line + Penetrator pwns attack groups until they get tactical missiles / sumos / goliaths, then adding in stumpys and bulldogs still counters them
- - Needs better artillery use (artillery operate differently from the main army)
- - Might want to add T1 gunships to follow attack groups to vary attack forms
- - Instead of small waves of attacks, make those waves less often and use them to scout at mid game, save all the excess troops for a rush and to defend extensions
- - Attack group size needs to increase with time and attack frequency needs to decrease with time

- No fusions/moho mine despite game lasting 1 hour (I was stalling to see)
- - Perhaps get the AI to save m/e for a T2 fac if territory expansion hits maximum
- - Seizing territory is less important once the economy hits T3, so getting enough territory to get to T3 while preventing your enemy from getting T3 as fast is the goal, not simple territory acquisition (this doesn't apply quite so much in CA, where metal maker equivalents are complicated and AI-unfriendly)

- Failure to get through a line of dragon's teeth = massive clog in the mountains which dies to 1 plasma battery
- - If units fail consistently to move through an area, list it as impassable and send scouts to test if it's become passable again.
- - Check for lots of wreckage and send a con with guard to reclaim a path if there's low enemy activity

- Died to a nuke despite me taking my time (1 hour)
- - A-Nukes required
- - Base is too damn small (got wiped in one nuke), needs to branch out when it starts getting clogged

- Air scouts penetrate way too easily. As it is, T2 gunships aren't countered at all, what little T1 AA just dies; 20 bombers will kill the AI commander
- - Basic air cover needed, even just an unassisted T1 airplant on repeat fighter set to patrol is better than ground-only AA (set land at 80% + 1 air repair pad at T2 is a bonus)
- - Failing that, needs to build T2 AA at least

- Massive lag towards end game, had to pause to give orders to my units

[romulous]

Got a crash, first one on the new version of spring / e323 3.18.1. I've not played Arctic Inferno before now.

[Zydox]

Another crash on Tabula v2... me against 3 bots...

[1v0ry_k1ng]

what aspect is currently being worked on?

better group merging and group size esculation over time would improve performance vs noobs + AI

the application of aforementioned defence algorithim would certainly help vs small raids vs the AI that currently do massive damage, but most of all I want to see the application of KAI style defences on chokepoints

[Zydox]

An XTA Crash report (AI logs and infolog)

[Error323]

what aspect is currently being worked on?

better group merging and group size esculation over time would improve performance vs noobs + AI

the application of aforementioned defence algorithim would certainly help vs small raids vs the AI that currently do massive damage, but most of all I want to see the application of KAI style defences on chokepoints

You can observe what is being worked on at github! Check the issues and commits.

[Zydox]

Isn't it time for a new version soonish? :)

[CerealKiller159]

If anyone wants to experiment with an up-to-date build (1 commit old as of now), this should help.

I compiled it myself for windows and it works fine for me.

http://www.mediafire.com/file/mzhzduf0gzt/
Extract to Spring\AI\Skirmish\E323AI\3.18.1\
You might need to make a new folder. Don't forget to copy all the other ai files from a previous version if you do.

(The included configs are optional, I just copied and renamed the other BA ones for BA 7.12, and added geothermal powerplants as emakers, allowing the ai to build them)

I hope, at least, that this helps prevent any redundant error reports of already-fixed bugs.

[Error323]

Well there you go it seems . I'm currently working on some more sophisticated profiling to see where optimization should occur. And I want nano turrets!

[Zydox]

Faster teching would be nice as well... as it is now, it seems like the bots will just spam units into your defence line instead of teching up and attacking with L2...

It would also be nice to see some L2 mexes in BA :)

[Error323]

Have you tried playing with the config file? Also I guess it heavily depends on the map. If there is a lot of metal (e.g. comet catcher redux) it'll tech faster.

[Zydox]

Have you tried playing with the config file? Also I guess it heavily depends on the map. If there is a lot of metal (e.g. comet catcher redux) it'll tech faster.

I tried poking around a little, though from what I saw there wasn't much to change in the config file...

I lowered the max techlevel which made them tech faster, but I didn't find any way to make them stop feeding the players metal...

I'm not sure if it's possible to create something which keeps trac on how the attacks has gone, so that if the bots conclude that they are charging into static defences without killing anything... they will stand back and tech up or better units instead of just feeding metal to the enemy

[Error323]

Zydox,

Could you perhaps record and post a demo of the behavior you are referring to? Also state the ingame times (begin:end) when this occurs so I can get a better understanding of what you mean.

[Zydox]

Zydox,

Could you perhaps record and post a demo of the behavior you are referring to? Also state the ingame times (begin:end) when this occurs so I can get a better understanding of what you mean.

Sure... here's a replay... I'd say somewhere around 13min ingame, the bots should switch to L2 instead of feeding me metal...

Also noticed that the replay doesn't show what happend in the battle... at the end my base was full with flashes... in the real game they were attacking the bots, though I guess that's an issue with the engine and not the bot...

MP battles against the AI also seem to have a tendency to desync...

And this battle crashed...

[hoijui]

Skirmish AIs (except Lua ones) are not able to desync, except through bugs in the engine, or if they alter the FPU state (which i am pretty sure, E323AI does not do).

last weekend, i had some AI MP test battles too. the other guy was running spring 0.81.2.0, i had 0.81.2.1, and we got hard desyncs. it should not happen, and it might be cache related or something, but try to make sure you all have the same version, next time, and report back if this fixed the desyncs.

[AF]

Skirmish AIs onyl run on one machine, so they dont have anything to desync with. Their internal execution is internal and seperate from the simulation. Their involvement in the simulation is purely via the commands that they give out just like any other player.

Thus if a native AI can desync a game, then so must a human brain.

[Zydox]

Crash logs for E323AI 3.19.0

Logs & Demo (8mb)

[Error323]

E323AI v3.19.0 - High Templar
Changelog:

• builders now also repair
• implemented factory pattern
• solved crash in CGroup.cpp
• don't build on mexspots taken by allies
• better profiler, outputs to "timings.dat"

E323AI SO/DLL: HighTemplar
E323AI BUGS: Issues
E323AI SRC: Github

And we're back! I fixed the latest crash issue and no more crashes found so far (tested 5 games). Furthermore the build now released contains a DEBUG level 2 build which makes the spring infolog.txt very usefull as we can now see where it crashed in the code. So please add it together with log and demofile when stuff goes wrong.

- Error323 & Slogic