Loading...
  OR  Zero-K Name:    Password:   

ZK Graph-Based AI (ZKGBAI)

242 posts, 11167 views
Post comment
Filter:    Player:  
Page of 13 (242 records)
sort
Skasi
8 years ago
quote:
It now builds pylons somewhat earlier and with higher priority so that grids are built much faster and more reliably.

That's worrying.
+0 / -0
edit: nevermind, that was server problems.

this version is definately improved, 1v1 it reliably beats down circuit so long as circuit dosn't reach striders first - awesome job :)

When will support for large team games be added? At the moment it dosn't appear to consider allied units and builds towers next to allied towers etc?
+0 / -0

8 years ago
The AI interface does not have good support for teams atm. It's difficult even to detect if the ai is playing on teams or not, and there are no events for things like allyCreated, allyDestroyed etc so no ability to defend, repair or whatever as far as allies are concerned. And yes, it also builds n times redundant porc for each kgb in the game. :| I'll probably start hacking at the ai interface upstream soonish, but it'll be a while before any of that hits ZK. (assuming I get all that done before spring 101 is released) Even then teams is complicated for lots of reasons, including resource constraints.

The lag spike issue is real but I'm not really sure what causes it. It seems to be related to units retreating or something along those lines, and it's fairly rare and not easily tracked down.
+0 / -0

8 years ago
how do you debug anyway? not that i got time to contribute with my adult life (which makes me sad)...
+0 / -0


8 years ago
Before array threatmap we had a nice debug window that showed all maps that the AI had, so you could use this and your usual error logging / debugprint :)

I'll get the window going this week i think.
+0 / -0

8 years ago
AllyEvent stuff has been mostly/completely implemented. :P Once I've confirmed that there's nothing else that I need to do I'll make a pull request and hopefully it'll be in the next engine, whenever that may be.
+2 / -0
8 years ago
Observation based on a few games:

ZKGBAI even at +60 metal/second does not seem to build anything interesting like a strider or a superweapon, and just keeps spamming small units?

Seems pretty boring.
+0 / -0


8 years ago
I've seen it go striders plenty, but by design the AI tries to finish the game fast rather than escalate and Eco like circuit does- it's a 1v1 AI atm not a team games one.
+1 / -0
8 years ago
ZKGBAI 0.6
the AI does no do anything. in zK playerlist it shows as <waiting> and if it is the only player on its team then it is instantly game over. If there is another AI on its team then only the other AI spawns a com and plays.

[f=0000000] Skirmish AI "Bot1" (ID:0, Short-Name:"CircuitAI", Version:"0.7.9.2") took over control of team 0
[f=0000000] Skirmish AI "Bot3" (ID:2, Short-Name:"CAI", Version:"<not-versioned>") took over control of team 2
...
[f=0000000] Skirmish AI "Bot2" (ID:1), which controlled team 1 is now dead

http://pastebin.com/ZcCRnS6b
+0 / -0
Hosting latest zkgbai (0.6.0) and latest circuit (0.7.9.2) together crashes spring at endgame screen, 100% reproducible (5 out of 5 attempts) on GNU/Linux. No backtrace in infolog is present, only message from lobby in console that spring has exited with code 11 (which means segfault).
+1 / -0
quote:
[f=0000000] Error: Initializing AI Interface library from file "G:\Spiele\SpringLobby\engine\100.0\AI\Interfaces\Java\0.1\AIInterface.dll". The call to initStatic() returned unsuccessfuly.

Something broke your Java AI interface.

Of interest: i'm observing exactly this failure mode on Mac build of Spring. Possibly it is, thus, not exclusively Mac-related.
+0 / -0
quote:
Hosting latest zkgbai (0.6.0) and latest circuit (0.7.9.2) together crashes spring at endgame screen

That's probably because of some magick Java interface failures. Try zkgbai and NullAI or Shard (both C++) or any other non-Lua AI: it should also crash.
Java is so heavy and troublesome nowdays. :P

But after major OS updates and recompiling engine with new ABI (gcc5 feature) i can't reproduce zkgbai's crash on linux. :)

WRT zkgbai 0.6: performance improvement is great, nice work.
+0 / -0
8 years ago
all my spring versions were downloaded with springlobby by joining autohosts
+0 / -0
What about your java?

Did you do anything peculiar to your env variables recently? Do you have a JAVA_HOME?
+0 / -0
8 years ago
Version 8 Update 66 (latest), just the user runtime thing, not the JDK, nothing configured, other java based software works
+0 / -0

8 years ago
I had a similar problem which seemed to disappear after installing both the 32 and 64 bit JRE versions.
+0 / -0

8 years ago
Ah that's probably it. I think on windows spring requires the 32 bit JVM, which most people would not think to install on a 64 bit machine.
+0 / -0


8 years ago
watching these replays..

http://zero-k.info/Forum/Thread/20792

I think ZKGBAI makes too much porc too early.. ability to naked expand where not threatened would help a lot.
+0 / -0

8 years ago
There's no such thing as "not threatened". All it takes is one flea, glaive, dart etc to kill an undefended mex. Also, the official version of kgb does not have any sense about teams, and by default builds defenses before it builds a mex. That basically leads to slow expansion, double porc and other inefficient nonsense. The current dev version knows not to build porc before mexes on teams, but still builds way too many defenders, doesn't team build big energy or anything else, and generally is not efficient in teams. The AI interface also doesn't provide events for allies so kgb cannot defend its allies when they're attacked, nor send AA when its allies are attacked by air, which are central parts of kgb's current strength. As a result, team games are not very interesting for testing KGB's behavior or making design decisions, at least for now, and probably until spring 101.0 is released.
+0 / -0
Skasi
quote:
but still builds way too many defenders

Wow I didn't think that was a possibility. :P
+1 / -0
Page of 13 (242 records)