M28AI Devlog (v149)

[maudlin27]

@kagurazakanana
Thanks for the feedback, in respect of your points:

• AIx Overwhelm issues - I'm assuming you're playing with AIx M28, could you send me the replay ID where this doesn't work? Testing just now it's working as I'd expect.

• 40km 3v3 - If you send me the replay ID I can have a look at what it did for if there are any obvious improvements that could be made. I suspect given the map size and it being a naval map it was simply too slow to capitalise on its higher AIx modifier, since it may choose to invest heavily in navy (and I'm assuming the paragon was outside of naval range)

• Not attacking with T4 air - This will depend on how much anti-air there is, so I'd need a replay to check whether it was being over-cautious

• Bomber snipes - it should intercept strats with asfs (as well as having the ACU under a shield, with health upgrades, and SAMs). My guess would be this was on a 40km map and its air force was too far from the ACU when the strats were spotted (but if it had asfs and didn't even try to intercept the strats then send me the replay)

[kagurazakanana]

Hi Maudlin27,

Thanks for your prompt response! Here's my feedback and follow-up on the points you raised:

Replay Upload:

Unfortunately, our last game was played on a local LAN setup, so no replay was saved. I’ll make sure to upload the replay from our next match for you to review once we've played another game.

AIx Overwhelm Settings:

Regarding the Overwhelm Rate, Overwhelm Interval, and Overwhelm Limit, I did notice that M28’s economy improves over time, but the build rate (e.g., engineers' build speed) didn’t noticeably change. One specific example: M28 was set to a (1.21.0) multiplier, with Overwhelm values of 0.1, 5, and 10. By the 28-minute mark, a T3 power generator should have been producing 25001.2(1.1^5) = 4831.53, but it was only producing 4000. Similarly, T3 mass extractors were only producing +28 mass/second when I expected them to be around +52.

40km Naval Map:

Yes, it was indeed a naval map, and it’s based on an older version you can find on the FAF client under the name "Durex 40km".

Paragon and Naval Battles:

The Paragon was outside of naval range, and the AI spent considerable resources on its navy. However, the AI couldn't effectively leverage its battleships. The players were able to focus fire with battleships and used nukes to clear specific targets, leading to a cost-ineffective navy for the AI. This allowed the players to cheaply fend off sea attacks.

T4 Air Attacks in Late Game:

M28’s air strategy was solid during the early and mid-game (with many successful gunship strikes), but by the late game, even though it had significant T4 air units, it stopped attacking aggressively. In that match, the defending player had about 300+ T3 sam AA, 500 ASF (with 50 T3 air factories constantly producing), as well as a large number of Aeon T2 move AA and cruisers; however, these defenses were spread thin across the map. Meanwhile, M28 had 10-15 CZARs, along with many gunships and ASFs, which could have broken through. I suspect M28 evaluates the total number of AA units rather than factoring in the map size and positioning when deciding whether or not to proceed with an attack.

Unit Sharing After Ally Defeats:

One observation was that once all of M28’s allies were defeated and their units were shared, M28 started attacking with those units, but by then it was too late. I suspect M28 isn’t accounting for the availability of allied units earlier in the match, which could have provided an advantage.

Bomber Snipes and ASF Response:

M28 did respond by sending ASFs to intercept once it detected T3 bombers, but my teammates deliberately distracted the ASFs by sending forward CZARs and Soul Rippers. This shifted the AI’s focus, allowing bomber attacks at the map’s edge to go unnoticed.

Misuse of ASFs:

In some cases, M28 ASFs would dive directly into dense anti-air zones, leaving behind large wreck fields. There were also times when the factory rally points were placed inside AA zones and remained there for the rest of the game.

Some Minor Suggestions for Improvement:
Allied Unit Management:

Multiple AI allies producing the same type of experimental units (like CZARs or Soul Rippers) don’t share them efficiently amongst each other. Perhaps consolidating some of these forces under the control of one stronger AI could make their use more effective.

Mass Extractor Sharing:

M28 AIs on the same team don’t seem to gift upgraded mass extractors to their allies. When one AI upgrades its mass points, sharing the upgraded extractors with allies could help boost their economies, especially on larger maps.

Engineer Sharing:

M28 AIs on the same team don’t share engineers, which could improve characteristic and construction efficiency.

Experimental Construction Prioritization:

M28 doesn’t seem to prioritize certain experimental units correctly. It often builds Salvation or Scathis, which are ineffective against shields, rather than opting for more impactful units like Paragon or Mavor.
-some m28and game setting:
1.5*1.5 (other def)
no mod open(only m28ai(v142) )
Thanks again for all your efforts. I’ll get you the replay from our next game as soon as possible for further review.

[maudlin27]

@kagurazakanana Thanks for the suggestions. Some are already present in M28AI, others the approach taken is intentional, while others I'd need a replay to consider further:

• AIx overwhelm - The main thing I'd be interested in is if the limit of the AIx overwhelm doesn't apply properly - I'm more relaxed if the timing is slightly out on when the bonus gets applied, as in the basic sandbox tests I've applied the bonus seems to take place roughly when I'd expect. So for example if you're applying AIx overwhelm with a 1.5 limit, and well after the time that it should've reached that limit it still isn't at +50% resources, then send me hte replay and I can check.
• Paragon and naval battles - replay would be needed - e.g. the one scenario that might be possible to adjust is where M28 starts losing significant navy to nukes, in which case it could switch to other units like a game-ender. It should already reduce battleship construction once it gets a large amount if it has control of the waters (so my guess is the battleships keep dying to things like nukes, causing it to keep investing in navy instead of a game-ender). A replay would simplalry help re deciding whether it's worth adding logic to switch to a game-ender where the enemy is getting one and it can't be reached with navy or non-amphibious land units.
• T4 air attacks - If there were 500 asfs then there's a good chance it'd be enough to dissuade M28. If it's getting to the point where M28 has a czar for every 25 enemy asfs and it's still not attacking then let me know and I can do some sandbox tests, but the enemy having c.50 asfs per czar doesn't sound too unreasonable for it to be cautious (from memory the break even point was somewhere between 30-50 asfs when I sandboxed a couple of years ago as a human player)
• Impact of allied units - yes M28 ignores allied units for the most part, with the exception of human players if shared armies is enabled. This is intentional, as particularly with weaker AI (like the default AI) M28 can't rely on those units attacking at the same time as it. It'd also have cpu and memory implications if it was to factor in those units the same as its own.
• Bomber snipes working due to experimental decoy - Although it evidently cost it the game, I'm relatively relaxed about such a decoy working, as it means the AI can be tricked, and often an experimental air unit presents its own existentional threat for the AI
• Misuse of ASFs - I'd need a replay for anything like that, it's meant to avoid SAMs in most cases (an exception would be if it had a czar or other high value unit nearby it was trying to protect)
• Multiple AI allies not sharing - M28 doesnt have this problem, all M28 AI on the same team will share all units between themselves (i.e. it applies a hive mind approach to control for the most part).

That said, where the team is spread out on a large map it might divie itself into 2 'hive-minds' for certain things like air units.

• Extractor sharing - M28 gifts mass storage to whoever owns the mex to ensure the adjacency bonus. It doesnt need to share mexes between other M28AI as it shares mass instead

• Engineer sharing - Engineers work together to build something, e.g. if you have 4 M28AI on the same team, and they all have engineers in the same small area, then they'll work together on big projects such as experimentals. Coupled with the resource sharing mentioned above, that means this would have minimal impact (the main use case, which M28 already caters for to some extent, is a teammate's engineer can block construction of a unit)

• Experimental construction prioritisation - It prefers paragon and mavor to scathis. Salvation gets built if it's eco so high already that paragon wouldn't provide a significant benefit. What I expect is happening is that the Cybran M28AI is in a position to build a game-ender, and has no T3 engineers of another faction near where it wants to build the game-ender, in which case it builds scathis. This is intended, as it's a balance between M28 going for what it htinks is best, allowing players to have influence on how the AI will play to some extent i.e. having a Cybran in a team so there's a chance scathis gets built), and avoiding scenarios where the Cybran in a safe eco slot doesnt build anything, while the UEF in the front 'mid' type slop gets a mavor (that is easily killed due to being on the frontline).

[maudlin27]

v144 Update
9 changes, including:

• Transports shouldn't drop a zone already being dropped
• Pre-emptive PD against further away enemy threats
• Asfs should be less likely to engage non-priority enemy air threats if lacking air control
• 2 LOUD specific changes - more Blackops ACU upgrade support, and M28's nuke dodging logic should now be supported in LOUD

Acknowledgements

• Vortex - several replays, and highlighting that M28's nuke dodging logic wasn't working

[maudlin27]

v145-146 Updates

• In shared armies mode, if there is a mix of M28AI and human players on the team, M28AI should only consider the eco of the M28AI players when deciding if it is overflowing mass (to cater for a game played between players who would control M28AI's units while letting M28 choose what to build)
• 4 changes relating to M28 trying to better deal with gunships with MAA (it should consider retreating/consolidating its MAA forces in some scenarios, as well as retreating more generally if faced with a gunship threat it cant handle)
• 3 changes primarily related to improving M28's build order on the ditch and similar maps with huge amounts of reclaim so it scales its power and build power faster
• Slightly reduced the tendancy to build t1 arti early on
• 1 LOUD specific change (relating to getting t1.5, 2.5 and 3.5 units a bit sooner)
• Very minor optimisation and a couple of minor profiling bugfixes

Acknowledgements

• Tryth - replay of a shared armies game where M28 overbuilt engineers due to human teammates overflowing mass
• Sladow - several replays against M28, including one where M28 AIx 1.4 was beaten (on the ditch)
• Vortex - LOUD replay and highlighting potential improvements

[kagurazakanana]

Hi maudlin27
Thank you for your detailed feedback. I have gone through your response, and I appreciate the clarification you have provided on various issues. I’ve also taken your suggestion to prepare a replay showcasing one specific issue observed with the M28 AIX Overwhelm rate.

Replay Details:
In the attached replay, you can observe the following anomaly:

AIX Overwhelm rate inconsistency: In the replay, many of the factories themselves fail to respond to the overwhelm multiplier properly. In the later stages of the game, there are also instances where production power seems to get applied without any observable incremental increase. Additionally, energy and mass extractors seem to receive a partial benefit, but the benefit applied is inconsistent and varies across different buildings.
I believe this has to do with how the multiplier is handled in relation to newly constructed buildings, which may not be adjusted correctly.

Reference to FACN Platform Fix:
About ten years ago, I was working on a similar AI system for the FACN platform (similar to FAF but based in China, forked from an older version of FAF). At that time, while facing a similar issue with adaptive AI not applying overwhelm and cheat buffs accurately to new buildings, I managed to implement a fix to resolve it on that platform.

Here is an example of how I approached it in Unit.lua under Unit = Class(moho.unit_methods):

lua

OnStopBeingBuilt/--------------Enable AIX Detection Thread
if ScenarioInfo.type ~= 'campaign' then
local AIBrain = self:GetAIBrain()
if string.lower(AIBrain.BrainType) ~= 'human' and self:GetAIBrain().CheatEnabled then
if ScenarioInfo.Options.AdpAI ~= 'off' then
self:ForkThread(self.CheckAutoAIbuff)
end
end
end
Additionally, I implemented a periodic check function to ensure proper multiplier adjustments:

lua

CheckAutoAIbuff = function(self)
if not self:IsDead() then
if ScenarioInfo.type ~= 'campaign' then
local AIBrain = self:GetAIBrain()
if ScenarioInfo.Options.AdpAI ~= 'off' then
local refreshTime = (tonumber(ScenarioInfo.Options.AdpAI) * 60)
if not nextRefresh then
nextRefresh = refreshTime
end
if string.lower(AIBrain.BrainType) ~= 'human' and self:GetAIBrain().CheatEnabled then
local timeSecs = GetGameTimeSeconds()
RemoveCheatBuffs(self)
if timeSecs < nextRefresh then
SetupAICheatS(AIBrain,self,AdpCheatMult,AdpBuildMult)
elseif timeSecs >= nextRefresh then
if (AdpCheatMult <= 1000) and (AdpBuildMult <= 1000) then
AdpCheatMult = AdpCheatMult + (AdpCheatMult/2)
AdpBuildMult = AdpBuildMult + (AdpBuildMult/2)
end
SetupAICheatS(AIBrain,self,AdpCheatMult,AdpBuildMult)
nextRefresh = nextRefresh + refresh time
end
end
end
end
self:ForkThread(self.AIXWaitUnit)
end
end
This was one of the approaches that I found effective for resolving the inconsistent multiplier application for newly constructed buildings. However, with the passing of time, I no longer recall all the intricate details of this solution and how specifically it was implemented, which makes it challenging to apply the same exact logic to the current FAF system.

FACN Lua.nx2 File:
To support your investigation, I will include a FACN lua.nx2 file for your reference. This might provide useful insights during the repair process. Unfortunately, I could not attach the entire file due to its size (over 1GB). In this version of the AI system, I had successfully addressed the issue of inaccurate multiplier application on new buildings. However, as mentioned earlier, some of the finer details of that implementation have since been lost, and I’ve found it difficult to map this solution onto the FAF environment.

Additional FACN Code (for Health Scaling):
Regarding AI scalability, I had implemented a setting in FACN that allows the AI to have X times the normal health. This feature, when combined with the proper adaptive AI system configuration, can allow the AI’s health to increase progressively over time. Below is the relevant code that implements the health multiplier, which can be adapted further for dynamic health growth.

Example from Unit.lua/SetupAICheat:

lua

if ScenarioInfo.Options.AIMultHP ~= "off" then
AISetMaxHealth = tonumber(ScenarioInfo.Options.AIMultHP)
unit:SetMaxHealth(unit:GetMaxHealth() * AISetMaxHealth)
if EntityCategoryContains(categories.COMMAND, unit) then
unit:SetHealth(unit, unit:GetMaxHealth())
end
unit.HealthM = tonumber(ScenarioInfo.Options.AIMultHP)
end
Supporting function in Unit.lua:

lua

function setupHealthMultiplier(unit, Mult)
unit:SetMaxHealth(unit:GetMaxHealth() * 10)
unit:SetHealth(unit, unit:GetMaxHealth())
end
This code ensures that AI units can initially have a health multiplier (X times their normal health). To achieve the goal of having dynamic health growth over time, you can combine this multiplier setup with a proper adaptive AI configuration that adjusts the multiplier periodically. The health multiplier would then increase as game time progresses, allowing AI units to scale effectively in longer or more challenging scenarios.

Regarding Other Mentioned Issues:
As for the other points I previously raised (such as the behavior of naval battles, air attacks, and misuse of ASFs), I haven’t encountered those issues again as of this response. Should any of those issues arise in future matches, I’ll make sure to record the replays for further analysis and share them with you for a better evaluation. I look forward to understanding how these can be improved further.

Replay Settings:
M28: 1.0*1.0

M28AIX Overwhelm rate: 0.1

M28AIX Overwhelm interval: 0.5

M28AIX Overwhelm limit: 10.0

Mods included:

M28AI v145
All Faction Quantum Gate V2.0 (created by jayTac)
Map:

3v3 Durex v10 Hotfix
lua(FACN) (2).zip 3v3_durex_v10.v0005.zip lastgame.zip

[maudlin27]

@kagurazakanana said in M28AI Devlog (v146):

3v3 Durex v10 Hotfix

Do you have a copy of the map you used? I can't find it on the FAF map vault

Edit: Although doing a sandbox test of AIx overwhelm to check if it was adjusting the build rate I realised the implementation was only adjusting resource generating units, so the build rate adjustment should be fixed in the next version. If you still run into issues after this then post a replay (and the map if it's not on the vault) with an example unit and timestamp illustrating the issue and I can look into further.

[maudlin27]

v147 Update
8 changes, mostly relating to minor bugfixes, including:

• Fixed a bug with AIx overwhelm not increasing the build rate of units
• M28Land should consider building non-land experimentals occasionally late-game to give it a small bit of variety
• Fixed a couple of bugs that could result in air factories not building units in rare cases

Acknowledgements

• Fearghal – replay which contained a number of niche errors
• Kagurazakanana – Highlighting that M28’s Aix overwhelm modifier wasn’t adjusting build rates
• c04spoon - LOUD replay where M28 overextended

[maudlin27]

v148 Update
13 changes, including:

• Naval AA units should be more likely to retreat from enemy units
• Fixed a bug where naval units would think they were always outranged by the enemy
• Subs should prioritise shield boats and then other subs
• Low ranged skirmisher units should be more aggressive when lacking intel if the ACU is nearby
• 4 MOD specific changes, including support for mods that require a build rate for factory units, building PD if it can be built on water in LOUD, and having M28 make use of the Blackops experimental aerial fortress

Acknowledgements

• c04spoon - Naval replay prompting a number of this update's changes
• Vortex - Replay where M28 failed to support his ACU with land units

[maudlin27]

v149 Update
27 changes, including:

• New performance mode option in game settings which uses simpler approaches for parts of M28 logic to make it run slightly faster, although also not quite a strong. As a rough guide, I'd expect this makes things run 10-20% faster overall, and if combined with M28Easy and the maximum delay between land unit orders, it ends up being roughly 50% faster.
• 6 mod specific changes, including better support for experimental factories and the XtremeWars experimental spaceship (which should no longer idle), and more aggressive use of Seraphim bombardment ships when navy is won
• Increased experimental unit variety if M28's unit prioritisation is disabled (meaning it will build a much wider range of units, but they are much more likely to not be appropriate in the circumstances)
• Gunships should be more aggressive when they have a large threat and sufficient AA to handle the enemy asfs
• Reduced the number of LABs built early on
• Limited the number of engineers to be built on small islands

Acknowledgements

• Vortex - LOUD replay highlgihting issues with the Seraphim bombardment ship and overbuilding engis on islands
• WesMaggs - describing a scenario where M28's gunships were too cautious
• Radde - A number of 1v1s on Xander Adaptive, whigh highlighted some early-game issues M28 had, as well as some other niche issues (such as land experimentals being overly cautious in some scenarios)