M28AI Devlog (v148)
-
@thewatcheral said in M28AI Devlog (v59):
21423105
Thanks, looks like there's an incompatibility with the experimental shield mod that triggered when the experimental shield was killed (and is highly likely to have caused the crash), I'll see if I can resolve in the next release.
-
M maudlin27 referenced this topic on
-
M maudlin27 referenced this topic on
-
v60 Update
Various changes to M28 including a complete rework of how it shields high value targets, the introduction of a very simplified ACU snipe option (don't expect M27 levels of aggression for this), and other improvements that for example mean M28's probably favoured to beat M27 on setons now (one of the few large maps that M27 had left).- 4 changes from testing M28 against M27 on Burial Mounds (mostly relating to a greater focus on gunships to defend against an experimental in some e.g. low power situations)
- 26 Changes from testing some high mass/AiX modifier scenarios, although I gave up on trying to avoid M28 overflowing mass in its build order at very high modifiers (engineer pathing issues and delays with construction starting and unit roll-off times proved too great a hurdle to overcome)
- 5 changes to improve M28's novax defence logic, including fixing some issues that meant it was only building T3 shields instead of sometimes getting T2 shields
- New shield cycling logic for high value units (e.g. game enders and T3 arti) - prompted from a replay where M28 failed to build 1 game-ender in the time it took the enemy to get 4 T3 arti, despite M28 having double the mass income, which I think was in part caused by almost all of M28's resources being spent on shield assistance. This change involved a lot of work and although I've made a number of adjustments to it from testing, I expect further refinements will be needed over time.
- 5 Changes for misc issues spotted while testing the new shield cycling logic (adjusting bomber attack logic, T2 arti construction logic, and a bug on transfer of certain HQs on an M28 teammate's death)
- 17 Changes from fighting M27 on Setons, including greater ecoing, and various refinements to M28's logic when a paragon is constructed to make greater use of the income (e.g. aircraft carriers will start to build air units)
- 8 Changes for Aeon M5 and M6, including fixing an issue with Aeon M5 where the T3 arti were effectively invisible to M28 (it must have been by chance that it managed to kill them when I tested previously)
- 2 changes for replay 21516542 to reduce the tendancy of retreating MMLs and T3 mobile arti of running into the attacking enemy (if caught out of position)
- 11 Changes for various replays from Radde, including more detailed logic for determining where to build T2 PD (that I hope means it is less likely for T2 PD to be built where they'll just shot into a cliff), increase in land experimental aggression vs ravager, and making it more likely M28 rushes an experimental earlier on
- New simplified ACU snipe logic - it's not really working as well as I'd hoped (and is a far cry from M27's version) but I figured it provides a starting point that I can refine as/when I see it not working properly in future replays
- 5 changes from Sladow replays, mostly relating to early game build order points (more engineers should be built for expansion, land scouts shoudln't be requested for core bases)
- 4 changes from Watheral replay containing a crash, improving compatibility with the experimental shield mod (which could cause a crash due to a flaw in M28's logic), and allowing the Aeon Artemis Satelite Control centre to build and use satellites (previously it was an expensive paperweight)
- 2 changes for a game vs M27 on Open Wonder to improve M28's initial build order/stop it power stalling (M28 still loses to M27, but it's no longer as crushing a loss)
- 5 other changes, including no longer self destructing drones spawned by blackops/similar mods when close to the unit cap, and making M28's ACUs more likely to get the RAS upgrade
Acknowledgements
- Balthazar - Pointing me towards a blank blueprint template for use when tyring to find a custom sized area to build in, and allowing me the use of said blank template (this is used by M28 to identify locations large enough to support its new shield cycling logic)
- Jip - Adding a more detailed 'pathto' function to take into account blocking cliffs (allowing M28 a hopefully less cpu intensive approach to identifying PD placement locations than M27's approach)
- Radde - various replays highlighting flaws with M28 (in addition to the trophy winning replays) and thoughts on improvements
- Zhanghm18 - Identifying an incompatibility with mods that spawn drone units
- Sladow - Several replays and suggestions for improvements to M28's approach to the game
- TheWatheral - Replays containing crashes
Trophy awards
- Radde claims the Sladow trophy for beating v59 M28 at 1.7 AiX (winter duel)
- Radde, Mhad and GhillieWolf claim the Radde trophy for beating M28 with a 1.5 resource mod on mapgen.
Note for future Sladow trophy awards these will only be for 5km or larger maps (i.e. maps that are smaller than 5km x 5km like winter duel won't qualify) - essentially I've had to admit defeat on an AiX 1.5 M28 being able to win against a competent player on this map since unless I devoted a large amount of time to a very map specific build order and approach it is too easy to overcome the resource/build rate deficit. I've also changed the trophy/medal picture in the original post of the thread.
I may well at some point change it to be a 10km only (since that's the size Sladow beat m28 on originally) but I expect there will still be some improvements that can be made on 5km maps more generally even if they're likely to also feature similar issues for M28 as winter duel did.
-
v61 Mini update
Small update to fix some issues with the special shielding logic added in v60- 6 Changes mostly relating to making the ACU prioritise attacking enemies instead of reclaiming or building mexes if it has a gun upgrade
- 4 Fixes and redundancies relating to the special shielding logic introduced in v60 to try and reduce the tendency of M28's engineers getting stuck in a loop of building a unit only to reclaim it.
Acknowledgements
- Zhanghm18 - Reproducing a bug where M28 would build and reclaim the same unit
-
v62 Update
M28 should play noticeably differently on 5km and some 10km maps 1v1 following this update:- Introduced new 't1 spam' mode, where M28 will try really hard to avoid getting a T2 mex so it can invest the mass into T1 spam or factory upgrades instead. It'll need some refinement but it at least doesn't seem like a significant drop in competitiveness for AI games
18 misc changes, including:
- making 'short range' land units more likely to attack the enemy instead of waiting for the enemy to die to longer range units (which would in some cases lead to a GC staying back so that a lone MML could try and kill the enemy T2 PD)
- Adjusting logic for getting engineers to cancel their orders to reclaim so they'll try and resume their order once done if they were heading towards a mex, and making it less likely an engineer follows up 1 reclaim order with another for reclaim far away
- Improved M28's logic for thinking it has mass to spare in high AiX scenarios (e.g. previously if all the M28 players on a team combined had 4k mass income per second they'd ignore most low mass conditions which could lead to aircraft carriers spamming torp bombers)
- Added basic build logic for the tempest
Acknowledgements
- Zhanghm18 – Testing different scenario’s impact on memory usage
- Sladow – Tips on how to approach a 5km game (suggesting to stay at T1 the entire game)
Trophies
No change to the current holders from their defeat of v59 M28: Radde (Sladow trophy); and Radde, Mhad and GhillieWolf (Radde trophy) -
I'm seeing an issue where the AI can't decide what type of shield to build. It will queue up one type and immediately replace it with another type indefinitely. This keeps a number of engineers occupied (50-100?) I watched this happen for at least 10 minutes.
These are both tier 4 shields, but from different mod packs. Nuclear Repulsor Shields, and Wyvern Battlepack. Wyvern is no longer in the vault, but generally works fine.
-
@bsee Thanks, do you have a replay ID and timestamp? Also do the shields from those mods have a larger footprint than normal shields?
-
@maudlin27 The Wyvern shield does have a larger footprint.
I suspect you can't load the replay since Wyvern isn't in the vault anymore. However the replay is https://replay.faforever.com/21588120 and the issue starts at 21:45 at the very left edge of the map.
Another small thing is that engineers tasked to the naval factories spend a lot of time driving to switch between the assisted factories. They might be more productive if they stayed on their current factory.
-
@maudlin27 You can try this replay: https://replay.faforever.com/21595213
We updated Wyvern to v5, which is uploaded here: https://forum.faforever.com/topic/362/wyvern-battle-pack-v5-for-all-game-versions
At 18 minutes the Cybran AI tries building an experimental surrounded by shields between the 2 AI bases. But it's alternating between experimental types. For this match the Wyvern T4 shield was disabled, but it was alternating between village & city shields which also have a size difference.
-
@bsee Thanks, I expect the size difference is likely to be what is causing problems, I'll add a fix in the next release to see if it resolves it
-
v63 Update
Main changes this update are completing my working through of the SC campaign missions with M28 as a hostile / allied AI, and the introduction of a new AiX option for if using M28 to take over the campaign AI (where the AI will receive a modifier to its build rate and/or resource rate).- 4 Bugfixes, including relating to the use of mods that add shield units
- 11 Campaign related changes mostly relating to error messages that would appear from testing on SC Cybran missions 1-6
Acknowledgements
- Zhanghm18 and bsee - Noting issues with M28's new shield template logic when custom shield mod units are used
(No new trophy awards)
-
@maudlin27 I'm still seeing the issue. We disabled all shields with a differently sized footprint. The engineer is toggling between ED5, ED6, a village shield, and a cybran rapid fire.
Here's the log from the host: game_21602429.log
You can see in this replay at 15:20: https://replay.faforever.com/21602429
All the mods are in the vault. Look just north of the pink AI base. -
@bsee Thanks, the previous version fixed the issue with the different sized shields but there are a couple of separate issues still present reviewing your replay (the first I hadn't realised would happen is that enginers can end up not building anything if they have multiple T3 shields available to build, the second is that it can try and build experimental level shields instead of a scathis). Both will be fixed in the next release.
-
v64 Update
- Improved how M28 acts as a teammate with non-M28 players, so it should be much less likely to build all over a teammate's base
- 5 Fixes/adjustments to M28's late-game shielding logic, including fixing compatibility with certain mods that add additional units
- Make M28 less likely to attack land scouts with its ACU early-game
- 2 Changes based on a Sladow replay (M28 should be more likely to send an engineer to high value reclaim in a zone ahead of smaller reclaim, and should be less likely to overbuild AA buildings in response to bombers)
- 9 other tweaks and fixes
Acknowledgements
- Fearghal - Replay and highlighting M28 building on a (non-M28) teammate's base
- Bsee - Replay and highlighting M28's issues with certain unit mods late-game
- HintHunter - Replay showing M28 messing up its build order due to a land scout
- Sladow - Replay and tips for M28's 1v1 5km approach
-
@maudlin27 thanks for the shielding logic fix! Now we're seeing mid to late game a large group of engineers set to assist a single one, which has no build orders. This results in 100+ engineers not doing anything.
Two replays: https://replay.faforever.com/21622277 and https://replay.faforever.com/21622977
-
@ender06 Unfortunately both replays desync really early on for me so I can't reproduce. Having said that, it may not be a bug, especially if M28 is low on mass/energy or has run out of places to build, as it tries to assign some engineers to assist shields (increasing the amounts based on how many arti and novax the enemy team has)
-
This might be a better replay (fewer mods): https://replay.faforever.com/21630530
By 19 minutes you can see M28 has plenty of mass & power, but most engineer are idle. I see a paragon queued up but for some reason the engineer just isn't starting it.Unrelated note: Does it actively avoid building saucers? It seems to build every other T4 but I haven't seen a saucer in a long time.
-
@bsee Thanks, one issue is if it tries building its late-game template near the map edge - I'll have it fixed in the next release.
The other issue is I'm guessing a FA game engine bug where a unit can be given an order to move and build, yet doesn't move - a human player would just give a slightly different order, whereas the AI tries the same order again and again. I've added a workaround for the scenario in that replay, which hopefully will resolve most instances (as I already had a workaround present for the main engineer logic, but my 'shield template' logic is a special case).Re Czars there are lots of reasons why it might not build them. Generally, I'd expect it to try and build them if it has air control, hasn't moved into the 'build T3 arti/game ender' stage (which is more likely at the high mass income levels in your game), hasn't built lots of experimentals already, and has space to build them. It is intentionally less likely to build them than Seraphim is to build Ahwassas though (mainly because Ahwassas are far better than Czars, even moreso for an AI)
-
v65 Update
A mixture of bugfixes, small adjustments, and a bit more personality for M28:- 6 Bugfixes, including with M28 trying to build its special shield template on the edge of a map (and failing to build half of it), and fixing an issue where M28 would think it was defeated if a non-M28 teammate died
- 4 Early-game adjustments, such as delaying when M28 upgrades to T2 when it decides it wants to spam T1 land
- 5 Changes based on replays from Sladow, including fixing an oversight with land scout production that meant they would be far less likely to be built
- Added a wide variety of messages for when M28 is killed or wins the game
Acknowledgements
- Bsee replay highlighting an issue with M28’s late-game shielding where the location is near the map edge
- Sladow replays and noting issues such as M28’s lack of land scouts
- !MarLo – confirming that I likely can’t get Campaign videos/audios to play due to how FAF loads campaign files.
Trophies
No change to the current trophy holders (although Sladow came close to claiming the trophy bearing his name) -
v66 Update
Hotfix for a number of bugs identified in a replay, with 10 fixes and other changes made, including:- Fixed several bugs causing engineers to idle late-game
- Fixed major bug introduced in v65 to M28's asf/intie logic that would cause them to stop moving
- Fixed an oversight with M28's scouting logic that would effectively cause spy planes/air scouts to get stuck 'scouting' a mountain ridge and never scout the enemy base
- Adjustments to what experimental gets built for Aeon, hopefully improving the decision making after a paragon is built and making a Czar more likely to be built
- MAA should be more cautious when the enemy has a stronger nearby land force.
Acknowledgements
- Lancer Dudo – Noting they fought against M28 when posting their replay for review; the replay highlighted a number of bugs with M28, one of which (its air scouting logic) has presumably been present since the original release of M28.
Trophies
No changes to the current trophy holders from v59: Radde (Sladow trophy); and Radde, Mhad and GhillieWolf (Radde trophy). -
v67 Update
Another small hotfix update, involving 10 fixes/other changes, including:- Fixed an issue with the Holy Raid custom campaign mission where an M28 teammate would reclaim the Cybran labs you were meant to protect (causing the mission to fail)
- Fixed a bug with M28's decision on what experimental to build (introduced in v66) that could cause its logic to fail in some cases
- Added a further workaround for nuke launchers to try and reduce cases where a yolona oss ends up idling forever
- Slight increase in start and end of game messages
Acknowledgements
- QAI3000 – Replay highlighting bug with M28 reclaiming objectives on the Holy Raid campaign mission.