AI Development Guide and M27AI v74 Devlog

maudlin27

v33 Hotfix

Fixed a number of Plateau related bugs that resulted in issues where the Plateau had 3+ mexes
Fixed bug causing bombers (in particular T1 bombers) to not work properly when in defence mode
Various other bug fixes (23 in total including the above 2 points), including Czar targeting issues, MAA not being built for experimentals, and SMDs not building more than 2 missiles.
36 misc changes, including making it less likely M27 will overflow mass, and improvements to special T1 bomber engi and mex hunting logic.
Misc changes included tweaks to the logic for building PD to make the AI build more of them (originally intended as part of a larger update to have the AI make firebases in some rare situations, but this part was postponed due to the number of bugs that needed fixing).

Thanks to:

Pyrolol – Highlighting issues relating to Plateaus on Open Palms
Relent0r – Highlighting a bug with the bomber defence logic and Plateau issues on Open Palms, along with replays showing a Czar targeting bug, and bug with sending MAA with experimental units
Fearghal – Replay highlighting SMD not unpausing after firing its missile (when 2+ missiles had been built)

maudlin27

v34 - Dynamic firebases

This update introduces firebases, although the scenarios where they get built will be relatively rare. More extensive turtling logic is planned for a future update (specifically pre-planned firebases), but for now M27 has what I'm calling 'dynamic' firebases, which in practice are likely to be built if it faces a constant stream of progressively stronger enemy units. As part of this, it will now build T2 artillery for the first time.
Another related change is the introduction of shield assist logic to help with T3 arti wars, with M27 being significantly better at surviving against a T3 artillery.
15 Bug fixes, including an issue that stopped torpedo bombers working (they'd only attack enemies that got near them previously), and an error that would happen when building experimentals if the enemy had an experimental
33 misc changes, including improving ecoing ability slightly by trying to always have something upgrading (sometimes this can hurt it though)

Thanks to:

Pyrolol for highlighting a bug with Plateaus on Twin Rivers
Relent0r for a bug with the decision on what experimental should be built.

maudlin27

v35 - pre-planned firebases
Part 2 of my turtling logic - this time M27 will preemptively setup a firebase at map chokepoints. It only works on very simple chokepoints (e.g. Astro and Fields of Isis; there will be a number of other chokepoint maps which it doesnt detect/work on), and I'm not even sure if it is beneficial even on these maps (it has a slight boost in performance against AI but that's to be expected as most AI will suicide units into the firebase - for example it was sometimes able to beat DilliDalli AiX 2.0 on Astro Craters. It's eco performance suffers massively from spending so much on the firebase though)

Other changes:
24 Bugfixes, including

Novax and T3 arti targeting (Novax would go to a unit's original position; T3 arti could sometimes error out when selecting a target)
Fatboy run logic (that'd cause them to flee in terror from the slightest enemy threat)
Issues with the power-stall manager, in particular a miscalculation of the energy usage of SMD and SMLs
22 Misc changes, including:
Air factory builder tweaks so air scouts are more likely to be built, and so the air factory and intie should be rushed if enemy air units are detected early in the game
Making ACU more aggressive if the enemy doesn't have T3 units

Acknowledgements:

Jip – Replay that resulted in an error relating to a land factory being built on a plateau where the pathfinding didn’t work.
JantDarvus – replay highlighting M27’s vulnerability to early jesters
Relent0r – highlighting issue with fatboys retreating, and issue with Ythotha’s attack-moving (I’ve decided to park the Ythotha point as a potential future change as I think I want experimentals to attack most of the time, except when there’s a Ythotha death-ball nearby)
Spyro – Replay where M27 may have caused a hard-crash (It's not clear from analysing the replay if M27 caused the crash or if it was a separate occasional error that happens with FAF where M27 isnt active, but the replay highlighted a few errors/issues)

maudlin27

v36 - Skirmishers
Dedicated skirmisher logic introduced (as opposed to previous logic which would sometimes have some units kite):

Units well suited to skirmishing (hoplite, mongoose, sniperbots) should now be built
Skirmisher units should be much harder to kill by retreating from enemies that get closer to them at a much earlier point, and maintaining good intel coverage
8 bugfixes and 9 misc changes, including a new emergency T2 arti builder as a last resort for when enemy naval units get close to base, and making pre-emptive SMD get built sooner than before

maudlin27

v37 - Late game and game-enders

Added logic for building and using game-enders (Aeon will build rapid-fire arti instead of paragon since I expect building a poorly defended paragon in M27s base when it struggles to make use of all its mass without a paragon would do more harm than good)
M27 should now sometimes build quantum gateways and RAS SACUs
Tweaks and bugfixes relating to shields, so they're better at defending against enemy T3 arti (in particular Cybran)
Adjustments to bomber logic so T1 bombers are less likely to suicide into SAMs and more likely to prioritise units threatening mexes
Mass stall manager implemented
20 other misc changes and 12 bugfixes

Note that the game-ender logic is very niche, so in most games I wouldnt expect any to get built.

Acknowledgements

Azraeel – Replay highlighting M27 suiciding bombers into SAMs, and M27 Cybrans failing to shield T3 artillery
Fearghal – Noting M27 doesn’t work when its start position is underwater (for now M27 will just give a chat message warning about this, and I'll park this for if I ever introduce navy)
Unknown – Note that M27’s experimentals can run from enemy experimentals (causing them to die without doing any damage)

Edit: V38 Hotfix:

Fixed bug that would break M27 if any unit restrictions were enabled (thanks to Punzler for the replay highlighting this)

splitice

warning: Error running lua script: ...ed alliance\mods\m27ai\lua\ai\m27factoryoverseer.lua(524): attempt to perform arithmetic on field `?' (a nil value)
         stack traceback:
         	...ed alliance\mods\m27ai\lua\ai\m27factoryoverseer.lua(524): in function `DetermineWhatToBuild'
         	...ed alliance\mods\m27ai\lua\ai\m27factoryoverseer.lua(1219): in function <...ed alliance\mods\m27ai\lua\ai\m27factoryoverseer.lua:1107>

Finished in like 10 minutes flat after norush. AI did not build much.

maudlin27

@splitice Thanks for flagging, have you got a replay, and were you using the latest version at the time (v38)? I checked the line referred to in the log and couldn't see anything obvious, and the current WIP version (v39) works ok still on a test game with norush so I'd need a replay to narrow it down.

maudlin27

v39 Update - Niche units: Quantum Optics

The first update focusing on niche units - Aeon can now build the Quantum Optics Facility. Main changes this update:

Quantum Optics will be used in some cases where intel is needed
Various tweaks to the bomber defence logic to tone it down slightly/reduce its tendency to suicide bombers
Various adjustments to the logic when losing air control
New T1 PD hunter logic to help when civilians have a T1 PD near our base (engineers should also now avoid T1 PD - previously they'd ignore it)
Unit cap logic to handle scenarios where the unit cap is reached
13 Bug fixes, the most noticeable are likely to be issues with skirmisher logic that would lead to them either running too early or not running at all, and a bug that caused M27 to stop building more land factories if it suffered from a powerstall in the past
37 other misc changes, including better support for multiple nukes, improved T3 arti targeting, and emergency T2 artillery builder enhancements

Acknowledgements:

Azraeel – Lots of replays of M27 vs human matches, which highlighted yet more T1 bomber issues, M27’s vulnerabilities when losing air control and/or faced with mass restorers, and issues when multiple nukes fire at a similar time
Jip – Highlighting error messages on open palms relating to identifying reclaim segments of interest.

tyne141

@maudlin27 - I like the AI and have had some good games against it. I enjoy the blog too. I did notice a small bug on Setons where it air-dropped engineers on the plateau but never built on the mass extractors. The engineers just sat there for most of the game with untouched mass extractors right beside them.

maudlin27

@tyne141 Thanks, have you got a replay ID, as I've not been able to reproduce on either the FAF or non-FAF setons?

Also was this the latest version (as v39 fixed a couple of bugs that in some scenarios would cause Plateau engineers to not build mexes)

splitice

@maudlin27 It was a modded game. It would probably be a pain to replay https://replay.faforever.com/17253648

I saw there was an update since I last updated so I gave that a shot and got no where. On start:

warning: false\000M27ERROR Count=1: GameTime 4: sBlueprintToBuild is nil, could happen e.g. if try and get sparky to build sxomething it cant - refer to log for more details
warning: stack traceback:
warning:         [C]: in function `assert'
warning:         [C]: in function `pcall'
warning:         ...nder forged alliance\mods\m27ai\lua\m27utilities.lua(74): in function `ErrorHandler'
warning:         ...d alliance\mods\m27ai\lua\ai\m27engineeroverseer.lua(3146): in function `BuildStructureAtLocation'
warning:         ...er forged alliance\mods\m27ai\lua\ai\m27overseer.lua(7173): in function <...er forged alliance\mods\m27ai\lua\ai\m27overseer.lua:7154>
info: BuildStructureAtLocation: Had category to build. oEngineer=ual00011; UC=1; All blueprints that satisfy the category={ table: 27460820 
info:   "drlk005",
info:   "uab0101",
info:   "uab0201",
info:   "uab0301",
info:   "ueb0101",
info:   "ueb0201",
info:   "ueb0301",
info:   "urb0101",
info:   "urb0201",
info:   "urb0301",
info:   "xrl0002",
info:   "xrl0003",
info:   "xrl0004",
info:   "xrl0005",
info:   "xsb0101",
info:   "xsb0201",
info:   "xsb0301",
info:   "zab9501",
info:   "zab9601",
info:   "zeb9501",
info:   "zeb9601",
info:   "zrb9501",
info:   "zrb9601",
info:   "zsb9501",
info:   "zsb9601"
info: }

(Ditto Mods)
https://replay.faforever.com/17253862

maudlin27

@splitice Thanks for the replays,

Re the second replay, the issue is compatibility with one of your mods. M27 will currently only apply engineer logic to land engineers, but after doing some debugging it looks like the mod combination you're using results in the normal engineer blueprint (e.g. uel0105 for UEF T1 engineer) being modified to contain the air category instead, which I'm presuming is the Flying Engineers mod. I'll add in a fix in the next version so M27 will try and build stuff/use the engineers where it detects this mod or can't identify any T1 land engineers that are buildable, but it's likely there will still be some unintended sideeffects/interactions.

Edit: Checking the first replay it's the same issue as the second (with the flying engineer mod causing issues) - this will be fixed when v40 releases

tyne141

@maudlin27 said in AI Development Guide and M27AI v39 Devlog:

nks, have you got a replay ID, as I've not been able to reproduce on either the FAF or non-FAF setons?
Also was this the latest version (as v39 fixed a couple of bugs that in some scenarios would cause Plateau engineers to not build mexes)

It wasn't on v39; I think it was on v37 or v38 as I played the game about 1 day before v39 released. The replay URL is: https://replay.faforever.com/17259833

It's a bit of a weird match as I was just testing M27 against Sorian out of curiosity so I was only human player in the match. I haven't seen the bug happen when I've used M27 with humans but I haven't played on many maps with plateaus with humans yet.

maudlin27

@tyne141 The issue might've been fixed in the last update but to be safe I've added some further logic to the (WIP) v40 to make sure that specific scenario shouldnt happen (the issue being the engineers were stuck assisting the land factory, and never had their orders reassessed once the land factory had finished building all the units wanted for the plateau)

tyne141

Thanks for the quick feedback; it's really great to watch the AI develop as you add to it. I'll try running a similar setup again on v40.

One other thing that a friend and I noticed when we played 2 humans VS. 2 M27s on V39 last night was that one time a commander killed itself by standing in the channel on High Noon and getting hit by 3 x T1 torpedo launchers. The commander seemed to retreat most of the time when it was damaged a certain amount but it didn't seem to detect it this time. The other thing was that the AI didn't seem to respond very quickly to T2 UEF cruisers attacking base buildings; it was still very challenging to win but we managed in the end partly through weakening their economy through cruiser missiles. I think we played about 3 games in a row and eventually won the last game. I can find the links for the replays if they're useful; I assume it was on V39 as I think we updated last night before playing. It's still really challenging as the AI was very quick to hit us with nukes and/or experimentals if we let it grow for too long.

maudlin27

@tyne141 Yes replays of games between M27 and humans are always useful even if there’s no obvious issue/error.

The t1 torp launcher/acu point sounds like a bug (Edit: Reviewing the code for this there was a flaw that likely explains what you've described which will be fixed in v40) while the vulnerability to cruisers is more of a weakness with its current logic that I need to try and improve at some point (so a replay that highlights this would be helpful).

splitice

@maudlin27 thanks very much.

I thought it might be mod compatibility but I was thinking Auto Reclaim. I look forward to seeing what the next release can do.

tyne141

@maudlin27 - here are the replays:

From memory, we lost the first two games but it was in the second game that one of the commanders died early because it didn't move away from the torpedoes; the remaining team still beat us convincingly by teching up and hitting us with experimentals. I think we won in the third game mainly through long range naval.

For context, we haven't been playing too long but we're easily beating the standard AIs and Sorian etc; we don't really play against other humans much yet so we're not exactly great players but it's offered us more challenge than Sorian or LOUD which are the other AIs we've played against so far.

maudlin27

@tyne141 Thanks, the replays highlighted a number of flaws/tweaks to make so hopefully it can give a bit more of a challenge in the next release.

tyne141

@maudlin27 - awesome, it's great to be playing against an AI as it is evolving so we know that we can't just learn how it behaves once and remember how to beat it.

Out of interest, is the intention to definitely add navy/sea eventually if there is time to do it well? Or is it potentially easier to make an AI difficult to beat if it concentrates its resources on land and air? I notice that some of the AIs that do build navy don't defend them too well so it is often easy to knock out naval factories in which case it is a massive waste of mass for the AI but I assume that an AI that can strategically use certain naval units could be quite tough to beat.