Test 1: Fields of Isis, AIx SAI (1.6.7b) Cybran vs AIx adaptive (duncane's 2.3) Aeon, 1.25 build / 1.5 resources, omni off, supremacy mode.
In the early 20 min they're both struggling near the middle without any clear advantage from either side. SAI spammed more T1 units, claimed more territories and had a slight score lead at T1 stage but was pushed back when the game reaches T2 and T3. SAI didn't spend too much in T2 or T3 (although it has lots of T3 factories) before starting to build T4 but up to this point they were still comparable in overall strength.
What kills SAI was the late game. Although there seems to be some improvements in late game build power, SAI continues to fall behind in the T4 race versus duncane's adaptive AI. They started building T4 at around the same time but SAI was obviously still lacking build power.
This is what you see for at least 50% of the time in SAI 1.6.7b:
This is what the adaptive AI did at the same time:
Adaptive AI also has the nuke out much quicker than SAI:
So it's clear that which one was gonna win. In fact the adaptive AI finished the game ender (salvation) in about 55 minutes, which was quite comparable to what human players would do. The SAI only finished a T3 arty in around the same time because it only had 1 or 2 engines working on it.
With that said, the SAI does have something better than the adaptive AI. The way the adaptive AI uses its experimentals was very ineffective. It's Czars wanders in front of the enemy base, chasing troops or attacking outlying structures while being shot by SAM. It'd be much more effective to do a comm snipe or killing reactors, even if it fail it'll destroy part of the base with the fall. It's GC stop in front of the base and tried to take out shields one by one, it refused to shot at other stuff before destroying the shields but there were so many of them. Things like T3 arty or TML should target shields first, but GC or Czar shouldn't. They should just go to the core of the base, chase the commander, destroy stuff in it's path.
SAI also used air better than adaptive. They're quick and generally responsive to threats. Unfortunately there wasn't enough of them (lacked grouping) to make a difference.
A human player could have finished the game with the first 2 GC:
In comparison, SAI did much better in using it's T4. In fact if this was an assassination game the SAI would have won:
End game stats, you can see SAI losing it's edge during late game:
11 experimentals + 1 game ender (salvation) + 3 nuke missiles
vs
4 experimentals + 1 T3 arty
|||Test2: In order to make sure what I observed wasn't due to faction or starting point differences, I repeated the game with the same settings except letting SAI control the aeon and adaptive control the cybran.This time the adaptive AI did better in the early game. It holds a slight advantage from the beginning, kept it, and gradually expended it until it won. However, like last time, the outcome wasn't clear until the game reaches T4.
The adaptive AI uses a special combination of units, looks odd but sometimes quite useful---hordes of LABs running towards enemy firebases to attract fire, giving T3 units behind them more time to deal damage:
Too many LAB wasn't a good thing if the unit count was low or if they meet an enemy experimental. In this test SAI managed to get 400+ kills on a GC before it dies.
Adaptive AI was also using lots of T2 ghetto gunships. However the T2 transport acted like a T1 transport: it never carries more than 6 LABs. I guess duncane needs to do a minor adjustment here.
Adaptive AI continues to show it's superior building power. The next screenshot seemed a bit too ambitious at 24min into the game but the AI didn't stall and all 3 were finished in a timely manner. One of the reason being that it has 1.5x resources, another reason was that the adaptive AI constantly adjust it's assisting engineers by assigning them to one of the "priority" projects.
Game ender completed in 42min, that was impressive:
I guess that's what the SAI should try to imitate in terms of late game development.
End game stats:
|||wow, that's an interesting read. i think i was aware of much of it already, but having it compiled in one neat package made it much more clear.hopefully this will lead to improvements in both SAI and AIF.|||Test 3: The third test was performed using AIx SAI Rush and AIx Rush. The goal was to test other aspects of AI performance while minimizing the impact of T4. Both sides were UEF so that they don't have a useful T4 that could decide the game too early. (Yes, fatboys were useless in AI vs AI games)
Syrtis Major, 1.25build/1.5resources, omni off, supremacy
The SAI Rush shined in this map with it's inhuman ability to use TMLs. Although that's not enough to singlehandedly win the battle, it did severely hampered mass extraction activities of the adaptive AI, giving SAI an economy advantage throughout the game.
At T1 to early T2 stage SAI was clearly better in doing what it's supposed to do: rush
The rush AI was still holding it's ground, and when they both teched up things started to slow down. It is harder to advance in higher tier combat (except T4) because they take longer to kill each other and they move slower. This map also has lots of hills and slopes that caused aiming problems. A stalemate started to form as chock points gets filled with debris.
Would it be possible to make them move on if they've stayed on a target for too long? (maybe exclude artillery units)
The rush AI also has an unhealthy fetish in LABs:
It's worse on this map because lots of blue units (rush AI) were stuck, wasn't moving at all during late game, most of them mech marines. They just stand there and become ready-to-eat veterancy food for long range artilleries.
Despite having a weaker economy and lots of stuck units, the Rush AI was still doing OK in building stuff, at least it doesn't seem to be falling behind too much. It did this with a better system of assisting engineers:
Number of engineers in SAI rush base
Number of enigneers in AIx rush base
AIx rush assisting
Sometimes the SAI also shows good assisting behavior, but it's not as consistent as AIx rush:
That, however, could be a good thing in this test. SAI just wasted less mass on something that will turn into a wreckage after moving 10 meters. But if they're not playing UEF, that'd be a different story.
Anyway, it's interesting that they nuked each other at the same time. Yellow(SAI) has an antinuke and it's first missile was completed and fired ~0.3 second before the enemy nuke landed, too late.
After that, Rush AI continue to waste mass on fatboy and gets bombarded by TMLs. SAI finished a Mavor, and it took another 45min for it to kill off the rush AI. (would be faster if that's assasination)
I was too tired and forgot to capture the end game stats, sorry.
Quote:|||hey superdog,
I was testing it, too. I also found out that the AI plays better on certain map positions, then on others. If you switch the sides of two AI and test it again, it often happens that the one AI which was doing well in the first battle, looses because of this. Their expension bases are also based on the starting position, which may be good on some maps, and bad on others. Guess that's also a reason why they play on certain positions good or bad.
vashan|||Yes I realized that. That's why I did test 2 and the AI's behavioral pattern and general competitiveness were relatively consistent. That means even if certain starting position or faction diversity made a difference, that difference was still smaller than the difference between the two AI.|||Still working on the engineer stuff. Have some ideas for my next version.|||Thanks for test Superdog.... I hear ya
- Needs less LABS at later tech levels. This is a recent issue cause of the GhettoGunship stuff I added that took out the limit on LABS.
- Experimentals need to target better. Should be easy to tweak.
- Improve defense against TMLs. This is hard, I think I have the main base TMD covered, but how do players do it for outlying mass points and units? Theres no mobile TMD. Also Sorian has put in some great code that launches TMs ahead of units so it can even hit mobile units pretty well which even a human has trouble with.|||- Experimentals need to target better. Should be easy to tweak.
In regards to this I think the issue may be scouting, it not seeing the buildsings. I like the order T4 attack currently. Although this may be because I play assassination. Which do you think?|||Good to hear some responses from both of you. :)
Quote:|||I am not going to do t2 GGs at all... so I will reduce lab construction at that tech level
As for the experimental targeting - I have checked it and it does go for the commander first. Here's the list:
'COMMAND',
'EXPERIMENTAL ENERGYPRODUCTION STRUCTURE',
'EXPERIMENTAL LAND',
'TECH3 ENERGYPRODUCTION STRUCTURE',
'TECH2 ENERGYPRODUCTION STRUCTURE',
'TECH3 MASSEXTRACTION STRUCTURE',
'TECH3 INTELLIGENCE STRUCTURE',
'TECH2 INTELLIGENCE STRUCTURE',
'TECH1 INTELLIGENCE STRUCTURE',
'TECH3 SHIELD STRUCTURE',
'TECH2 SHIELD STRUCTURE',
'TECH2 MASSEXTRACTION STRUCTURE',
'TECH3 FACTORY LAND STRUCTURE',
'TECH3 FACTORY AIR STRUCTURE',
'TECH3 FACTORY NAVAL STRUCTURE',
'TECH2 FACTORY LAND STRUCTURE',
'TECH2 FACTORY AIR STRUCTURE',
'TECH2 FACTORY NAVAL STRUCTURE',
'TECH1 FACTORY LAND STRUCTURE',
'TECH1 FACTORY AIR STRUCTURE',
'TECH1 FACTORY NAVAL STRUCTURE',
'TECH1 MASSEXTRACTION STRUCTURE',
'TECH3 STRUCTURE',
'TECH2 STRUCTURE',
'TECH1 STRUCTURE',
'TECH3 MOBILE LAND',
'TECH2 MOBILE LAND',
'TECH1 MOBILE LAND'
There is a check that if a shield is protected the target it will shoot the shield first so that maybe what you are seeing.
Also if it sees the commander it override its current target and will attack the ACU first.|||Quote:|||would it be possible to have the GC (and also ML, and CB) stomp about and crush instead of just zapping things? after all, they have enough health to walk all the way through the shields and onto them.|||BulletMagnet|||1500 damage per footfall event, sounds pretty viable. i assume that it would still shoot while moving.
No comments:
Post a Comment