M28AI Devlog (v130)

In regards to the rules, it likely wouldn't, but this is more academic (UI mod is done, in fact, just in moderation limbo). I couldn't get it (automated tactical launches) working consistently with the stock SimCallbacks, and couldn't get it working consistently with my own custom SimCallback. I got curious enough to try to implement everything on the sim side; possible I missed something in the process, but couldn't get it working there either.

I've bandied it about a couple of times on the discord but no leads--thought you might have been one of the few to have discovered something given that it's a niche topic and you'd have cause to interact with it. If I find anything out I'll pass it back your way.

Hello!

I think the AI you created is very excellent, and it may even have a score of 1300+ in a 1.2*1.2 2v2 scenario. However, it's not perfect, and I have found several issues with M28ai. I started a 2v2 game on the IS map, where I produced a large number of ASFs in the late game, and our side hoarded some land T4 for defense. I then started building the game terminator, but the M28ai I was up against did not respond to my actions.

After I hoarded some ASFs, M28AI often attacked my T3AA position with small groups of troops. These failed attacks prevented M28 from hoarding effective ASFs.

After I started building the game terminator, M28, as an opponent, did not respond to my actions. This includes TML/bomber/some teleport.

I noticed that the AI prioritizes maintaining shields, but this is not effective. The AI did not repair enough shield arrays when dealing with T3 heavy artillery and mavor, resulting in heavy artillery easily destroying antinuke/game terminator facilities.

At the end of the game, when the player finished building the game terminator, M28 AI had an absolute advantage in land force (AI 8-10land t4 vs player 6 T4). However, it did not attack the player's base but kept making detours on the map, which resulted in the ACU and base being completely defeated.

replay.zip

@kagurazakanana Thanks, there's a map specific issue with the pathfinding that was causing the asfs to suicide, I'll see if I can have it fixed in the next version.
Not sure at this stage though re the game-ender solution - M28's current playstyle (to give a bit more of a differentiation to M27) is to typically try and win without using T3 arti, but on a chokepoint map like Isis that means it's possible to defend against any force it can build up. I'll have a think about a tweak to make it slightly more aggressive where it has a chance of pushing but I don't expect it to have a significant impact.

v20 Update

• 13 misc changes, including improving buid order where the hydro is relatively far away from base, making the ACU slightly more aggressive early-mid game, Ahwassa targeting tweak, and special shielding bugfix
• Improved ACU combat performance on FA Mission 2 (Dawn) so M28 can now handle the opening section on hard (if supported by a competent human)
• 21 changes to try and have M28 complete FA Mission 1 on hard - it managed to save the civlians (2nd main mission) once, but most of the time will still fail. This includes torpedo bomber micro, naval micro where units are outranged, fixing bugs with building naval factories in water zones, reducing the engineers sent to a water zone in one go once it has a factory in place, fixing issues with land scouts trying to move to locations outside the playable area, improved ACU build order in scenarios where it is overflowing mass due to reclaim, and increased gunship aggression
• New cruiser bombardment logic, where UEF and Seraphim cruisers should be more likely to target enemy structures instead of providing AA defence
• 3 changes relating to Fields of isis replay (including fixing a bug with path generation and making units more aggressive if the enemy gets a game ender)

Acknowledgements

• Ice-IX – Highlighting flaws with M28’s hydro build order on crag dunes.
• bsee replay where M28 failed with its map setup logic
• Hdt80bro – discussion about ways of checking if a unit has fired the last shot in a cycle
• Kagurazakana – Replay highlighting bug in M28’s land zone generation logic that could lead to asfs suiciding into the enemy base

@maudlin27 Did something big change in M28 AI in last days, hard to judge from version updates ? I played a few 10's of games on the dark heart. Usualy 2v2v2 with me and M28 vs 2x M27 vs 2x RNG. And while M28 was quite ok, it was still a lot less dangerous than M27. I thought that was normal behaviour that M27 is stronger than M28.
But now i played 2 games on eye of the storm, me and M28 vs 2x M27. M28 was bottom left, i was bottom right, each of us handled 1 side. He managed to kill the top M27 before i did, from the looks of it he completely dominated him, his score is insane, 3 times higher than mine and M27's. Did M28 somehow get crazy better in last 7 days or M27 got worse ? Or its just a lucky pick with the map ? This was really impressive from M28. If ur interested, replay id is #20557680.

@goblinsly2 Yes I accidently broke M27 when it's played with M28 in the last M27 update, it'll be fixed in the next version
(it's not an obvious 'break' as M27 still functions and makes a lot of similar decisions to before, but it's competing with M28 for who gives its units orders)

v21 Update
A failed regicide attempt by M28 - I figured I'd see how it performed on a selection of maps against M27 (I've held off until now due to how demotivating it was during M28's pre-release to see how heavily it got crushed by M27):

• Strip Mine - M28 win - 4 changes made, reducing the amount of MAA built generally, and adjusting asf engagement logic
• Astro craters - M28 win - no changes made
• Theta passage - Mostly M27 wins - 4 changes made (making ACU prefer combat over reclaim, reducing the amount of T2 PD to be built, and adjusting the attack logic of T1 arti when retreating)
• Polar depression - M27 win - 24 changes, which still werne't enough to yield a win. Includes improvements to initial build order, making ACU more aggressive, reducing tendency to pre-emptively build shields and SMD, greater focus on ecoing at the T3 stage (this change may make M28 worse!), and tweaks to retreat logic to try and reduce the tendency of units to block retreating sniperbots.
• M27 also won on Setons and Burial Mounds, although M28 managed to equal or better M27's eco so it wasn't a crushing defeat.

@maudlin27 Hello! I really like your work. Of all the AI played so far, your AI is the only one that can launch 5 unit mods and play on a 20km map for more than 30 minutes without crashing. When playing other ai, the game will pause at about 25min and then crash.

But I do want to reflect on one issue: UEF's AI will normally build almost all MOD units. But Cybran's AI basically doesn't build experimental units in mods. What I mean by this is that UEF's AI randomly builds a lot of the various experimental units that are added in the mod. Cybran, on the other hand, may only pick one or two of these experimental units to build during an hour of play. And cybran's AI has very little desire to build experimental units, he just keeps churning out various land armies and air forces, not even the Navy, until he runs out of 1000 unit caps.

Perhaps the AI can have a simple logic: let him reserve 50-100 unit cap, and then randomly build experimental units or navy.

@zhanghm18 The desire to build an experimental unit is almost identical between the factions, so I'm guessing it was different game-states that led to UEF building lots and Cybran not.

The choice of what experimental to build will vary though. The logic is based on unit categories, and if the desired category isn't available then the AI will switch to building a random experimental. What that means in practice is that if for example you have a mod that adds a bunch of land experimentals to the game, and M28 wants to build a land experimental, it will pick from any of the existing FAF and new modded experimentals.
If instead M28 wants to build say a novax and the mod removes novax but adds a new air experimental, then M28 will switch to building any random experimental (which will include the modded experimental as an option).

In terms of the unit cap M28 should already adapt to getting near the unit cap, e.g. it should ctrl-K T1-T2 units, not build more T1-T2 units (with a couple of rare exceptions), and stop building a particular type of unit that it has lots of (regardless of tech level). If it is actually hitting the 1k unit cap consistently I'd need a replay to assess, as adding logic to not build any T3 units at a certain point could have significant drawbacks - for example, if it stopped building asfs it could quickly lose air control and then it wouldn't matter how many land experimentals it built if they'd just die to enemy gunships/strats.

@maudlin27 I ran a lot of tests, on different maps. But the cybran AI will always build a large number of T3 land armies and a small number of air forces. And only a handful of experimental units will be built. In contrast, his rival UEF ai builds a large number of random land experimental units. As a result, UEF's AI wins every time

@zhanghm18 I'd need a replay to assess, as adding logic to not stop building T3 units at a certain point could have significant drawbacks

v22 Update
This update revisited how M28 performs against RNG and DilliDalli following the various updates made to M28. Generally it performed notably better than the last time I tested on a range of matchups against these AI, although it still struggles to not get overwhelmed at the T1-T2 stage, particularly against DilliDalli.

48 Updates were made on a map by map basis:

• Cadmonium Green - 3 updates vs RNG (forgot to test for DD)
• Stickleback Ridge - 8 updates vs DilliDalli (no changes for RNG)
• Crash site - 3 updates vs RNG, 3 updates vs DilliDalli
• Forbidden pass - 7 updates vs RNG, 1 update vs DilliDalli
• Open palms - 4 updates vs RNG, 5 updates vs DilliDalli
• Burial Mounds - 3 updates vs RNG, auto-win vs DilliDalli (it suicides its ACU into civilian PD)
• Badlands - 2 updates vs RNG, 9 updates vs DilliDalli
  (also tested against RNG on Pelagial, Eye of the swarm, white fire, Setons, and theta passage)

As a summary of some of the changes:

• Increased tank aggression early-game, with tanks being more likely to move defensively to protect expanding engineers
• Various bugfixes to TML targeting logic
• Early ACU build order tweaks
• Adjusting decisions on when to upgrade factories
• Aeon should favour blazes early on instead of obsidians
• Greater use of PD to protect both core and minor land zones
• Adjustments to how transports load engineers to try and increase the speed slightly
• Slightly greater use of MAA from before where the enemy has no bomber/gunship threat but lots of inties/asfs
• Adjustments to T2 arti building logic (fixing a bug that could lead it to be overbuilt while also increasing how early they're built in response to a fatboy threat)
• Fixed bugs with torpedo bomber targeting logic

Acknowledgements

• Balthazar – pointing me towards the :GetCurrentLayer() to determine if an experimental that is in water can be targeted by torpedo bombers

v23 Update
This update mostly covered niche scenarios, including certain situations in M1 and M3 of the Seraphim campaign and an issue late game if M28 reaches the unit cap and loses some engineers. Map loading speed has been improved significantly for maps with large amounts of water.

• Reworking of pathing related logic for water zones at the start of the map - for a slighty loss in accuracy in calculating travel distances of about 5%, loading time was improved for maps with lots of water zones (on the Seraphim M1 campaign map the relevant logic's time reduced from 71s to under 2s, with overall map time only taking 15s as a result for that map)
• Seraphim mission 1 - 10 changes, mostly around making the ACU more aggressive at the start
• (Seraphim mission 2 - fix made separately to FAF's Nav's navigational mesh - I'm hoping this might feature in the upcoming development update III next week, in which case M28 should start working on this map)
• Seraphim mission 3 - 5 changes, mostly campaign specific (along with a bugfix that could lead to too much PD being built)
• Zhangm18 M28 mirror replay (featuring high AiX) - 3 changes, mostly relating to having engineers more likely to spend mass instead of assisting shields late-game, to reduce the likelihood M28 ends up overflowing mass when it hits the unit cap.
• 3 other minor changes/fixes

Acknowledgements

• Zhangm18 replay – highlighting how M28 could overflow mass if it hit the unit cap and its engineers were all stuck on shield assist duty.

v24 update
This update was focused on fine tuning on the map four-leaf clover, pitting it in 1v1s against RNG, DilliDalli and M27, with 39 changes made:

• 24 changes made fighting against RNG, including improved initial build order; improvement to the location expanding engineers go to initially, increased ACU aggression, slightly improved air scouting, and expanded logic when deciding whether to attack enemy bombers/gunships to be more accurate in assessing whether it is safe to engage them.
• 3 changes made against DilliDalli, including pausing production to get energy for overcharge, and gunships being used to defend the ACU as a high priority
• 7 changes made against M27, including fixing an issue where land units could go idle if the enemy retreated, and fiing a bug that was causing excessive factories to be built at 'expansion bases' on the same island
• 5 other changes, including adding slightly more variety to very late-game experimental choice (Cybran and Aeon are slightly more likely to build air experimentals, UEF should consider novaxes again if it already has a Mavor), reducing the extent to which T3 combat units get built in unit cap scenarios, and reworking when M28's brain gets setup so the FAF Navigational mesh gets calculated on the playable area instead of the entire map (for non-campaign maps). I.e. this should mean M28 no longer thinks the two southern bases on Dual Gap can attack each other by land.

Acknowledgements

• Jip – highlighting how to get the NavMesh to only consider the playable area on non-campaign maps
• Relent0r – highlighting M28’s sub-par performance on four-leaf clover

Hi,

We've encountered an issue yesterday with M28. The two M28 AIs would not do anything upon game start. Here is the relevant replay in case it could help: https://replay.faforever.com/20630807
Do note the same map with M27 instead of M28 was OK (everything else was identical).

Another issue we had, that might not be related to M28/M27 at all: all 3 games we ran yesterday (2 with M28, 1 with M27) were laggy with frequent micro-freezes (no apparent connectivity issues between players, same PCs as usual and same group of players as we usually do). Again, this might not be related to your AIs at all, but I thought this was worth mentioning.

Thanks again for your hard work, it is very much appreciated.

@v_k M27 is more likely to have micro freezes. Do you have a replay and rough timestamp of an M28 only game with noticeable micro freezes?

@maudlin27 Hi,
Sure! This one with M28 had noticeable micro-freezes during litterally the whole game; I have no specific timestamp, it was kinda hard to play from the very beginning to the end. I'm not sure this is something visible in the replay though.
https://replay.faforever.com/20630393

Thanks!

@v_k said in M28AI Devlog (v24):

20630393

Does the replay stutter for you? I was able to run it smoothly at +10 speed throughout with no obvious stuttering, so I'm assuming the issue was one of the players in the game (e.g. I recall occasionally having a game involving only humans that stutters/goes very slowly due to issues on one of the player's side).