M28AI Devlog (v302)

maudlin27

v265 Update
8 changes, including:

T1 arti outranged by enemy units in one location should consider attacking nearby locations where they outrange the enemy
4 improvements to selens so they dont block engineers building when cloaked, and cloak from enemy gunships
Experimentals should reassess attacking a zone if its threat becomes negligible by significantly more than when it originally decided to attack
Fixed a bug where a 1.2 resource 1.0 build rate M28 AIx would have its ACUs benefit from a 1.2 build rate on completing an upgrade

Acknowledgements

Radde - Various replays and highlighting issues with the 1.2 build rate for ACUs and with selens blocking construction

maudlin27

v266 Update
6 changes:

Slight increase in ACU aggression in late-mid game and if facing an outranged enemy ACU
Fixed some issues with the late-game arti template (reclaiming blocking units; building ravagers instead of triads; getting UEF quantum gateway for shield SACUs)
Fixed a bug with finding the best target for t1 arti drops

Acknowledgements

Radde - replay against v264 which featured a couple of the above bugs.

maudlin27

v267 Update

Fixed the t1 arti target selection bug I thought I’d fixed in the previous version
Added some redundancies to assign more build power to power instead of experimentals when M28 is stalling E
Reduced the minimum amount of t3 land needed to start on an experimental in QUIET
Ahwassa should be more aggressive if pushed back to its core base, and consider attacking enemies in adjacent zones.

Acknowledgements
Relentor screenshot and comment that M28 was building an experimental when stalling E

maudlin27

v268 Update
7 changes, including:

Reduced the distance for ASFs protecting bombers from enemy air units (when M28 lacks air control)
Made ACUs more aggressive when they outrange the enemy ACU (again), and adjusted logic so they ignore enemy T1 PD threat if they have gun
Tanks should be less likely to attack an unupgraded enemy ACU
Adjusted decision on where to build PD to hopefully fix a bug where it could get built too close to the enemy

Acknowledgements

Radde - Several replays and highlighting PD placement bug

maudlin27

v269 Update
8 changes, including:

Fixing some issues that meant outranged tanks would suicide into the enemy
Significantly reduced the number of engineers to assign to tasks when they've recently run from the enemy
Selens should no longer be paused to save energy
More asfs should be built if M28 has air control but not by much

Acknowledgements

Radde - several mapgen replays

maudlin27

v270 Update
12 changes, including:

ACUs shouldn't cancel RAS to get stealth (unless at risk of dying)
MMLs should dodge missiles reflected back onto them by loyalists
Fixed a bug preventing M28's unit target priorities from having effect
Added a workaround for an issue where a megalith wouldn't fire despite enemies being infront of it
Fixed a bug with strat bombers being stuck hover-bombing if they fired their AA gun at an interceptor
An air experimental should be even less likely to be built before a land experimental

Acknowledgements

Radde - Various replays and highlighting a bug with a strategic bomber
Pmh1 - Noting a bug in campaign mission 3 where M28 wouldn't build ahwassas as the campaign AI (partially fixed; for reference if anyone else has similar issues, killing the engineers that were meant to build the ahwassas should hopefully cause the map to expand)

Trophy Award

Radde and KuhwaKlimakleber - Awarded the Radde trophy for beating M28 1.4 on a 15km mapgen

maudlin27

v271 Update
11 changes, including:

Sniperbots should be less aggressive against MMLs if they have lost radar coverage of them
Seraphim subs should be less likely to surface to use their AA attack and should submerge sooner when facing a combined air and navy threat
The SMD/SML nearest completion should be assisted by engineers if there are multiple in a zone
ACUs should cancel their upgrade and run from an approaching experimental slightly sooner than before
Added a workaround so cloaked selens should be ignored by most of M28's land combat logic

Acknowledgements

Radde - various replays and highlighting an issue with cloaked selens

maudlin27

v272 Update
8 changes, including various adjustments to land factories to make it less likely T1-T2 factories build MMLs and T1 arti if M28 has access to T3 land, stopping sniperbots running from t3 mobile arti if there's friendly T2 arti firing at the mobile arti, and increasing the range of manual T2 arti targeting so it considers targeting enemy mobile shields in preference to the default firing logic.

maudlin27

v273 Update
7 changes based on a live replay I came across, including:

Fixed a bug that would break M28's land combat logic if it had nearby friendly T2 arti
Gunships far from the main gunship force should no longer reinforce if it would mean traveling over enemy AA inbetween
Fatboys should build more mobile shields if against a novax
MMLs should check a larger number of nearby enemies to assess if they should run from that enemy

maudlin27

v274 Update
6 updates mostly based on a Radde replay (turtling against M28):

Land factories should build fewer units if the enemy appears to be turtling
Reduced the T1 arti to be built by T2 land factories
Added a small chance of M28 getting T3 arti relatively early on (in place of its normal logic)
T2 shields covering SMD should be upgraded to T3 if the enemy has t3 arti/novax and a nuke

Acknowledgements

Radde - 1v1 replay

maudlin27

v275 Update

ACUs should consider ctrl-king to kill the enemy ACU
T1 arti/MML should occasionally be tasked with raiding areas separate to where M28's main army is
Fixed a couple of issues with land scouts that could lead to an area incorrectly thinking it didnt need them, or to having selens attacking enemies far away from them

maudlin27

v276 Update
Small update based on a random replay I came across:

Transports should consider dropping far away zones with only 1 mex if there's also lots of reclaim
Gunships should prioritise shooting at enemy engineers ahead of tanks
Land scouts should be more likely to be built by a factory if it hasn't built any before (even if the team has access to better land scouts)
ACUs should run if they are in range of enough TMLs to snipe them (and full share isn't enabled/they have no teammates left)

hulgarth

Hello
I am playing a lot with m28 on spice 1 mountains and with easy and adaptive m28 they refuse to build nuke def on my team. I have no toggle to ask them to priorities it. I know the enemy ai is getting more resources when we lose the middle but I would say even with scouting 9/10 games the m28 adaptive ai even with huge resource deposits will get nuked every game. It is honestly like I am playing with a handicapped version on my team. I have taken to building nuke def for them but it is destroying our games. If I could get them to build even one atleast the games would be far more enjoyable. I feel like it is a bug because consistently the enemy m28 will build nuke def by 20 mins or so in every base. Even when I take full map control my ai will ping my anti nuke and never once build a game ender while the enemy ai every game builds a stream of satellites art and nukes.

I would say most of my games end with a nuke to my ai. I am lower skill but it feels like I am playing with a handicapped version. I tried changing the personality, turning off helpful ai but nothing works. Unless I am missing a toggle in the game settings. I only really play this map so it I am missing something I would appreciate any ideas

maudlin27

@hulgarth The latest release (v277) tries to make M28 more likely to build air scouts to try and detect nukes a bit sooner; the other option you could try is to scout the nuke yourself (i.e. build spy planes and periodically send over likely locations for a nuke) - if M28 detects a nuke launcher it prioritises SMD so the issue will be that it doesn't know the SMD is there (and even then it should still build some to be prudent, but just as a much lower priority than if it's aware of a nuke, meaning it can be vulnerable to a rushed/stealth nuke).

maudlin27

v277 Update
14 changes, including:

Making it more likely air scouts and spy planes should be built to try and reduce the likelihood the enemy gets a nuke without M28 realising
Adjusted the decision on whether to attack with outranged units so that they should factor in nearby zones attacking/moving to the same location (so for example if M28 has 2 GCs that individaully are too weak to attack a firebase, but combined are strong enough, it should no longer change its mind about attacking the firebase just because one of the GCs crosses the border from one zone to the other despite them both remaining by each other)
If helpful teammates is disabled M28 shouldn't build mass storage by (non-M28) teammate's T2 mexes

Acknowledgements

Fearghal - Replay and highlighting issue where M28 built mexes by his mex and didn't gift them (due to helpful teammate being disabled)

maudlin27

v278 hotfix

fixed a bug with the recent logic for zones to coordinate attacks
fixed several bugs preventing selens cloaking in LOUD
added a new feature to request an M28 teammate builds an smd at a location - use a text marker ping and say ‘build smd’ / ‘smd’ / ‘get smd’ and M28 should try to build one if it has enough nearby units (@hulgarth )

hulgarth

@maudlin27 Thanks for the reply. That has been my plan but if I forget or get them late they get nuked. In almost all my games the ai wins the naval battle at first. I will try to do early scouting but it is just odd that the enemy builds them like clockwork but never my side. I can get an early nuke off at 20 mins vs the easy version and hit one.

I am unsure if its even possible but a setting to say always build atleast 1 nuke def would solve it for all my games.

thanks

maudlin27

v279 Update
10 minor tweaks and bugfixes, based on 4 1v1s against M27 (one for each faction) on a mapgen map Relent0r highlighted around 76 versions ago as one where M27 would typically beat M28. I was relieved to see that following those c.76 updates M28 won the matches (with its recent raiding logic appearing to be a significant factor in its wins), but on the way I noticed some issues and/or error messages (and it had some very close games).

ACUs should be less likely to get reclaim outside their range and more likely to go to an adjacent zone that already has enemies in it (over one that doesnt), while fixing a bug that meant it wasn't having a slight preference to stick with the zone it decided to travel to in the previous cycle.
Fixed 2 bugs relating to M28's logic for trying to scout certain units of interest (such as SMD)
Land units should be less likely to attack if the enemy has units in an adjacent zone that outrange them, but in such scenarios should instead be more likely to attack if there is an enemy mex nearby

Acknowledgements

Relent0r - providing the mapgen code (around v203) that resulted in M27 beating M28

maudlin27

v280 Update
Mini update fixing 6 error messages that appeared in an M28 vs M28 mirror match. Generally the errors related to niche logic being applied to a dead unit (so these fixes wouldn't be expected to impact things significantly as M28 would normally self-correct for this within a second).

maudlin27

v281 Update
Changes based on one of the M28Easy 1v1 vs M28 AI vs AI tournament games (where M28Easy lost from a 2:1 eco advantage), and some basic testing of high mex low unit cap scenarios to avoid M28 being stuck not building anything.

Slight increase in upgraded ACU aggression
2 changes to make it more likely for TMD and/or T2 engineers to be built sooner if the enemy has TML
Mass storage shouldn't be paused before land factories (even where there is a nearby enemy threat)
Multiple land HQ upgrades at once should be much less likely even in high mass scenarios
Limited M28's t1 spam mode in 1v1 once at least 100 t1 tanks have been built
Disabled building land scouts to kite with for M28Easy (since it doesnt kite with them so they die easily)
10 adjustments to unit cap to cover things like ctrl-king T3 mexes, and permitting building SACUs
2 minor bugfixes