Skirmish AI: E323AI 3.22.4

[hughperkins]

Hey E323, it seems I may owe you an apology: I was using the version of E323AI that one gets when one does some kind of a pull in Spring, but it seems that spring's main git repository is locked to a particular version of each AI?

In your case, it looks like I was using your AI from mid-August... perhaps the one from September will slaughter my AI, there's a decent chance.

So, I probably need to get the sources directly from your repository, and retry against that version.

For the Spring git repository, I've posted in the development git thread http://springrts.com/phpbb/viewtopic.ph ... 34#p377934 proposing that there is an automatic submodule update and commit each day, so that the Spring git submodules stay up to date with the source submodules. Unsure to what extent this will be considered reasonable or not.

Hugh

[Error323]

Hey Hugh,

Spring is "locked" to a specific version indeed, they don't want an autoupdater because of security issues (I've asked the exact same question).

In any case, E323AI is located in spring as a submodule. You can always change dir to it and it will behave like any normal git repository.

Still working on the pathfinder, actually on an algorithm for creating a much better graph used by A*. Hopefully I have something tonight! Here's a WIP in the attachement, as you can see it still suffers from an error (arrow going across the red).

[Error323]

UPDATE v2.12.0 Thu Sep 10 19:17:57 CEST 2009
Changelog:

• New graph algorithm
• Optimizations in the pathfinder

E323AI SO/DLL: E323AI-v2.12.0-Ultralisk
E323AI SRC: E323AI@github

Under linux it performs well, zero crashes so far tested on:

• Delta Siege Dry
• Crossing 4 final
• Karst
• Barren
• MoonQ20x
• Small Divide

Please report any crashes or bugs you find.

[FaerieWithBoots]

i really it hate to disapoint you. However, the AI shows some idle commanders in the mid game again. One commander kept his "wait" command after the lab he was guarding got destroyed. The other had no reason to be idling. I can send you the replay if you like to. The game was played on Karst with XTA9.585.

Also, it doesnt work with XTA 9.59. But that might not be a problem in the E323 ai.

Good thing is that it didnt crash on startup, like it used to with the previous release :)

[hughperkins]

Hi E323, I got round to compiling your git version from a couple of days ago, but the commander just did nothing I'll download the current git as of right now, the time of this message, and give that a whirl.

I'll attach the logfile to this message. It looks like something to do with calling Boost? Maybe it's already fixed in latest git, but just in case it isn't.

Hugh

[Argh]

Under linux it performs well, zero crashes so far tested on

What games / mods was it tested with? Just OTA clones? I haven't had an AI that would play any of the factions in P.U.R.E. for nearly a year without crashing, was wondering.

[Error323]

i really it hate to disapoint you. However, the AI shows some idle commanders in the mid game again. One commander kept his "wait" command after the lab he was guarding got destroyed. The other had no reason to be idling. I can send you the replay if you like to. The game was played on Karst with XTA9.585.

Also, it doesnt work with XTA 9.59. But that might not be a problem in the E323 ai.

Good thing is that it didnt crash on startup, like it used to with the previous release :)

Well actually I'm happy, no crashes and correct pathfinding from what ur saying. I'll look into the idle commander. :)

[Error323]

What games / mods was it tested with? Just OTA clones? I haven't had an AI that would play any of the factions in P.U.R.E. for nearly a year without crashing, was wondering.

Hey Argh,

If you wanna see if it works take the latest version and initialize it in P.U.R.E. Next check the log for wishlistpush errors. And update the config file for P.U.R.E accordingly.

[Error323]

Hi E323, I got round to compiling your git version from a couple of days ago, but the commander just did nothing I'll download the current git as of right now, the time of this message, and give that a whirl.

I'll attach the logfile to this message. It looks like something to do with calling Boost? Maybe it's already fixed in latest git, but just in case it isn't.

Hugh

Its very important to use the latest version =)

[1v0ry_k1ng]

this looks good, dling now. works with current spring, right?

[Error323]

yep

[Argh]

If you wanna see if it works take the latest version and initialize it in P.U.R.E. Next check the log for wishlistpush errors. And update the config file for P.U.R.E accordingly.

Will try it out.

[hughperkins]

Hi Error323,

Played two games against 2.12.0 this morning. Ok, you might say that is two days old now. And you'd be right. In any case

- first game: HughAI lost, because, well, the scouts didn't think twilights were interesting to kill, and the tanks thought that if there was only three of them, just sitting next to the commander instead of attacking lone twilights was a wise decision!

- second game, using a new version of HughAI. HughAI won

Replays:

http://manageddreams.com/hughaireplays/ ... 14.tar.bz2

Note that the start of the replays is blank for about 60 seconds or so as the AIs crank up, so you could just fastforward through that bit on x10 speed or something.

Hugh

Edit: corrected link

[mcgoo]

tested E323AI on Windows. Plays nice, but if i manage to survive a while the AI builds more and more bots and a huge part of them is just standing around. CPU-usage gets higher and higher and the game gets unplayable slow till it stucks completely

[Error323]

Hey Hugh,

You really got me scared there, but looking at the demo, it seems that E323AI isn't using the proper configfile delivered with v2.12.0 (It shouldn't build fleas at start). Then again this will only help so much, since ur AI is also agressive, good job. I need to work on defence now ;)

[Error323]

tested E323AI on Windows. Plays nice, but if i manage to survive a while the AI builds more and more bots and a huge part of them is just standing around. CPU-usage gets higher and higher and the game gets unplayable slow till it stucks completely

Hey Mcgoo,

Yeah this is a problem at the moment hehe, the system increases the group amount as time passes by. Working on that aswell =)

[hughperkins]

Ok... well, I just ran with what was in git. It's possible to include config files in git you know

What is the procedure to get the appropriate config file and install it?

Hugh

[Error323]

The appropriate config file is in git and in the tarball. Make sure to get this directory/file layout:

Code: Select all

fhuizing@jarvis ~/.spring/AI/Skirmish/E323AI/2.12.0 $ tree
.
|-- configs
|   |-- BA-categorization.cfg
|   `-- XT-categorization.cfg
`-- logs

And try again :)

[hughperkins]

Hi Error323, so I have a BA categorization, but not the other one. A diff with the one in your AI's git shows they're different.

I'll just copy them across manually for now. It would be nice if it was copied across by the make install in the future

In case it's useful, my build procedure was:
- in the spring git source, "ccmake .", setting all paths to the 'game' subdirectory, then 'make install'

.. running on Ubuntu Jaunty.

Hugh

[1v0ry_k1ng]

Ok, ive been playing with this a fair bit

firstly, mega congradulations, this AI kicks ass! beats all the other AIs and performs reasonably vs humans, and the raidering microing is incredibly cool!

however, the CPU usage is excessive- considerably higher than any other AI, and enough to lag my machine. Im guessing that is somthing to do with the huge amount of constant threat checks, since units seem to receive about 2 orders a second. might want to tone this down until the system usage is reasonable.

on the way it plays, it would hugely increase its performance if a greater priority was placed on lab assistance (a lab should be assisted even when near stalling really)

with defences, with both XTA and BA a quick and easy way to implement defences effectively is place an imperative to build one the cheapest defences (xta missle towers/BA llts) within 200 of metal extractors

AIs also play TA mods far better when they have build in build orders for labs (ie, instead of making random shit mix of units, they follow a sequence from a txt file. To hide this, you could have 5-10 diffrent lab build orders in the file so it varies between them)

the other thing that would greatly improve playing performance is to implement reclaiming- ie, for metal piles worth over 50, cons can have an action to reclaim them. increase the priority of reclaiming in proportion to the quantity of metal in wrecks and the distance of the wrecks from the con. in BA/XTA, have fark/necro performing purely as reclaimers.

anyhow, impressive stuff :D