M28AI Devlog (v301)
-
v282 Update
Looks like I forgot to post this at the time (around a month ago)!
12 changes, including:- ACUs should try and avoid blocking construction of a factory when retreating to their base
- Land units with a direct fire attack and longer ranged indirect fire attack should try kiting with their direct fire range
- Shorter ranged units should no longer be scared off by an enemy ACU when they have large threat (prompted by a scenario where 20 land experimentals tried kiting the enemy due to the presence of an ACU!)
- Expanded logic on coordinated land zone targets to use a similar approach when outranged units decide if they have enough threat to attack alongside any longer ranged units
- Idle engineers part of M28's game-ender templates should consider helping nearby construction (that isnt part of the game-ender logic)
Changes were based mostly on some random live replays I came across
-
v283-284 Update:
A somewhat delayed update as I spent a few weeks practicing for a recent 1v1 tournament (by playing lots of games against M28) which payed dividends with me being able to improve my previous 3rd place finish with....another 3rd place finish!34 changes (including 1 bugfix from a v384 hotfix):
- 4 ACU changes, including having injured ACUs run from destroyers, and M28Easy considering dodging T1 arti fire with just its ACU (since even weak players will be unlikely to leave their ACU stationery facing a bunch of T1 arti)
- 4 Air changes, including fixing an issue where on 7.5km and 15km and similar 'unusual' map sizes M28 could end up ignoring enemy air units, and fixing a bug that could lead M28 to rally its air units in range of enemy cruisers
- 6 Factory changes, including making M28Rush not build skirmisher units, M28Turtle and Tech building even fewer factories, and fixing a bug that meant bombers/gunships wouldn't be built to protect a friendly naval factory under attack
- 8 Engineer changes, including having M28 not try and build mexes in an area where a teammate has a base or forward position (previously they would consider as a low priority), to try and build structures away from the water if taking fire from a nearby enemy navy, and for engineers to be more likely to run onto land to get away from a naval threat (instead of staying in the water)
- 2 land combat changes, fixing a bug that could lead to an island not updating its request for reinforcements after killing all enemies on it
- 9 naval combat changes, including introducing logic to recognise multiple 'zones' attacking the same target at once (so their forces are combined when assessing if there are enough), including hover attacks from land, fixing some major issues with how the tempest was being used so it no longer prioritises ground-firing subs when enemy battleships/similar are in its range and so it considers kiting enemy naval units, and made it more likely units should try and attack a nearby exposed supporting/outranged naval unit if the enemy's supporting naval force is too far away to help
- Redundancy relating to potential messages M28 sends when it thinks it has won the game to try and avoid M28 gloating about winning a game when it was killed
Acknowledgements
- Relent0r - Replay where M28Easy suicided an ACU into T1 arti (by leaving it stationary)
- Radde - Various replays, with the main ones prompting changes being a 1.7/1.7 win against M28 on a naval map, and a 1.4/1.0 win against M28 on Selkie (naval map)
- Fearghal - screenshot showing M28 building a mex in a teammate's base on astro
Trophy award
Double trophy awards for Radde this time:- Radde trophy - awarded to Radde, Vortex7547, Mhad and kuritew for beating 1.4/1.0 M28 on a mapgen
- Sladow trophy - awarded to Radde for beating 1.7/1.7 (!) AIx on a mapgen
-
Hey @maudlin27,
first of all thank you very much for putting this much effort into the project.
Is there a specific reason why you let the AI build this many long range land units over direct fire units? I'm asking because this is not very realistic at all. If "not", maybe consider adding some sort of lobby setting drop down deciding about the unit mix (e.g. x% direct fire) would make it more realistic. Also maybe consider providing more options in the doding / hover drop down increasing numbers above three, maybe up to five or ten to have more levels than just three or inifinite.
To give you some context:
I do run a training group for beginners within the FAF Dojo three times a week (from 0 and playing the game pretty much for the first time up to 1200-ish rating), and we use the AI for team building (last game of the night all swarm the AI) and for training puposes to learn eco / tech if the peole that have showed up at the very day cannot be properly balanced in a PVP scenario.
-
@wilson_ I'd suggest having them fight against M28Easy or M28Rush if you dislike long ranged units (I presume you mean hoplites/mongoose/sniper bots)
-
Thank you for your quick response.
They're beyond the stage, fighting Easy or Rush won't help them evolve at this point.
I mean basically all long range stuff: mml, arty, the ones you mentioned above.
Is there a reason why you let the AI use them instead of using mostly direct fire and e.g. deciding when to go in via scouting? -
Because the AI has great apm, so can get more value out of long ranged units compared to a player (where the apm required to micro-manage such units means they're not as good for combat)
In the case of MML and arti, it will also depend on what the opponent builds (e.g. if enemy has sniperbots or PD it'll get more t3 mobile arti), and it seeks to have a small proportion of them in its unit mix both to help if the enemy gets PD, and to punish the enemy if they leave their tanks idle
-
just wondering, is there a difference other than version number obv, between the ai's in campaign/coop compared to in PVP? ive been using the bot in coop and it builds random stuff EVERY where, and like wilson stated it seems to over rely on mml, to the point of not being able to def its self. all of that is in coop ofc, which might be party why, since it could be responding to thing i cant see, or because there's actual objectives that it doesn't consider. but while watching old rainbow cup vods the ai's seem more or less... normal... not a big deal i was just curious.
-
@BLAAARGNN It's possible it doesnt play well on campaign due to changes made since I last tested it - I tested the campaigns using v91-96 of M28, and rely on bug reports for if it does something strange.
For example, a common reason it builds lots of MMLs is if the enemy has a firebase near it, in which case it responds aggressively with MMLs. I can't remember when the most recent changes were made with this behaviour, but it's quite possible there are campaign missions where it incorrectly thinks the enemy is getting a firebase (to get t2 arti within range of its core base) and so it spams MMLs to deal with it.
-
v285 Update
19 changes based on 2 replays from zwaffel- 6 engineer changes, including having engineers more likely to build away from water than before, build t2 arti to fight enemy navy, and naval factories are prioritised early if M28 cant reach the enemy with land
- 5 air changes, including building more torpedo bombers where lacking navy on a naval map, and increased hover-bombing micro on early bombers vs engineers
- 5 land combat and factory changes, including taking into account nearby enemy navy ranges when considering if M28's tanks outrange the enemy, being more likely to upgrade land factories to support factories, and reducing MMLs that get built in response to an enemy building threat in some scenarios
- 2 naval changes, including ignoring ACU's direct fire range if they're underwater
- 1 misc change, fixing an issue where if M28 first detects a unit such as an engineer on a beach that is neither on land, nor on 'proper' water, it wouldn't 'see' the unit - now it gets assigned to a nearby land zone (if there is one very close)
Trophy awards
Double trophy awards for Zwaffel:- Sladow trophy - Awarded to Zwaffel for beating 1.7 AIx (resource and build rate) M28 on a 5km mapgen
- Radde trophy - Awarded to Zwaffel for beating 1.5/1.0 M28 on a mapgen
-
v286 Update
4 land combat adjustments based on a rainbow cup replay on Fuji where M28 sent a megalith on its own towards enemy experimentals, then a short while later followed with a ythotha, with changes including:- Adjusted logic for when experimentals/high value units decide to manually attack a target to both add a redundancy to ensure they have visual of the target, and also allow them to manually target a unit if they already have an active manual attack order on that unit.
- Outranged units should attack the enemy if there is a friendly outranged experimental that isn’t retreating or can’t kite, that is in range of an enemy T3 land or experimental directfire unit.
For the curious, another change i looked into based on the cast of yesterday’s Rainbow Cup games was why tryth’s team’s pink M28 lost most of its base to a nuke on adaptive convergence. It turned out it was over 90% loaded with an smd, but as the enemy player team’s nuke wasn’t scouted until constructed and was heavily assisted, M28 couldn’t quite defend against it. The nuke proved decisive as M28 lost 200 asfs to the explosion handing air control to the enemy M28AI (a longer term adjustment I’ll be looking at is having M28 move its asfs if it can see a nuke missile approaching them, but I only had time for a small number of changes with this update due to wanting to release ahead of today’s RC; for similar reasons I wasn’t able to make M28Navy disregard hostile civilian ships to get its first naval factory setup, but it is noted down for some time in the future)
Acknowledgements:
- Fearghal – Rainbow cup IX (which provided the cast and replay prompting these changes)
-
v287 Update
9 changes, including:- M28Tech should eco slightly more than other AI (so it can be more suitable in eco slots on maps like astro and dual gap)
- Made some fixes and improvements to M28's logic for building units for raiding, which could lead to scenarios where far too many MMLs were getting built
- Decreased the chance of a first bomber for most M28AI (except M28Air and to a lesser extent M28Rush)
- Naval factory should be prioritised at the start of the game in more scenarios (including M28Navy personality)
Acknowledgements
- Radde - 2 replays with yew on mapgens
Trophy award
- Radde trophy - Awarded to Radde and Yew for beating M28 on a mapgen
-
v288 Update
13 changes:- 2 air changes including small numbers of torpedo bombers should ignore low AA threats (from 1-2 destroyers) on the basis after taking damage they can refuel
- 5 Engineer changes, including limiting how much T2 arti gets built if the enemy has a t2 arti firebase just outside their range, making M28Navy even more likely to build on the nearest pond to it, and refining whether engineers run from enemies (e.g. Aeon shoudlnt run from subs, and hostile civilians should be ignored if they might lack visual)
- 2 ACU changes - ACU should be more likely to cancel his upgrade even if in his main base (if he's on the outskirts of that base), with nearby enemy naval units also sometimes being included in the decision, and how quickly M28 is losing health (i.e. if it thinks it will be able to complete the upgrade before dying)
- 2 Naval changes, including naval units not suiciding into hostile civilian navy just due to having a nearby friendly factory
- 2 misc changes (changing naval factory pause priority, and not showing a chat message about being outnumbered if enemies are on different teams)
Acknowledgements
- Fearghal - Rainbow Cup IX, whose replays formed the basis of all the changes in this update
-
v289 Update
21 changes, almost all of which based on rainbow cup replays:- 5 Land combat changes including making fatboys less afraid of enemy ACUs that aren't in range of it and being more aggressive if it has significant supporting units to deal with enemy T1 spam, making MAAs run from destroyers in a nearby water zone, and fixing a bug that meant M28 was updating its last known position of a unit when it shouldnt
- 2 Naval changes, including a 1st cruiser should no longer be prioritised preemptively ahead of destroyers when the enemy has a nearby naval threat
- 11 air changes, including improved selection of air rally/support points so they are less likely to rally into enemy cruiser range, and fixed a bug that meant they wouldnt run from a sighted enemy nuke; Restorers should be much more aggressive against enemy air force if M28 has air control; and asfs should be more likely to protect friendly torpedo bombers
- 2 Engineer changes, including fixing a bug that meant Czars wouldn't be built by Aeon (which could also lead to a significant negative impact on sim speed)
- 1 misc minor bugfix
Acknowledgements
- Fearghal - Organising Rainbow Cup which prompted most of these changes
- Radde - Highlighting an issue with units knowing where an enemy ACU was despite lacking intel (due to a bug where M28's last known position of an enemy unit was being updated incorrectly in some cases)
-
v290 Update
9 changes, including:- 4 optimisations to reduce CPU load with large numbers of ahwassa and enemy MMLs
- 5 changes to make M28 prioritise reclaim early game more, including fixing a flaw in the code that meant if a rock near 1 engineer wasn't pathable all engineers in the zone would stop reclaiming (an issue that is particularly noticeable on Fuji Phantoms)
Acknowledgements
- Fearghal - Rainbow cup replays (in particular Fuji Phantoms game which prompted the reclaim changes)
- Radde - Sending a replay of a game with M28 (which wen I ran it went to -4 late game prompting the optimisations in this release)
-
v291 Hotfix
Updated to remove profiling that I forgot to turn off in v290 (and which I suspect may have been the cause of some intermittent crash reports) - unfortunately releasing the hotfix took some time due to a lack of internet.Thanks to Radde, Fearghal and zhanghm18 for reporting issues
-
v292 Update
9 changes, including:- Asfs should try and avoid rallying in-range of an enemy Czar
- Fixed an issue that could cause megaliths and other DF experimentals to keep firing into cliffs at a unit when taking T2 arti fire
- Fixed some M28 hostile campaign AI issues that could cause units to move off-map in various scenarios
- Added basic support for the CommonArmy game setting (so e.g. if you create a campaign mission with M28 teammates, and give M28 control of your ACU, it should function normally)
Acknowledgements
- Radde - replays highlighting different issues, including the above points re Czars and megalith shooting into a cliff
- Creed - Replay with CommonArmies campaign issue
-
v293 Update
15 changes, including:- Fixing a bug introduced in v292 which broke M28's 'hive mind' (causing every M28 player to act as though it was on a team of 1)
- Various changes to improve M28's ability to complete UEF missions 1 and 2, particularly if using M28's combined armies feature to give M28 control of the player. For example, on UEF Mission 2 M28 will now reinforce the civilians (including gifting the reinforcements to the player if M28 is just a teammate so the objective completes) and sending the civilian trucks to the relevant objective. Other notable fixes/improvements include fixing a bug that meant M28 would think a mex was upgrading and send its engineers to assist it, even if the mex wasn't upgrading due to unit restrictions, fixing a bug that could cause PD to be built somewhere non-buildable (likely primarily relevant for campaigns), and fixing an issue where if sharing control with M28 in combined armies and reaching the unit cap, M28 could get stuck trying to ctrl-k the same (non-shared) units when trying to manage what units to ctrl-k.
Acknowledgements
- Fearghal - Replay and log about significant CPU issues (and a memory crash) - caused by the 'hive-mind' failure bug; related issues (M28 underperforming) were also flagged by emerald, Santii and Danger
-
v294 Update
20 changes:- 2 further improvements to UEF M2 so gunships should be better at escorting civilian trucks to safety, and if the enemy ACU ventures out of his base they should consider sniping them.
- 6 improvements to UEF M3, including prioritising getting MAA at the start of the game and fixing an issue that meant M28 would try and capture arnold's black box (instead of reclaiming it).
- 7 improvements from a Radde Xander 1.2/1.0 game, covering scenarios where M28 has 2 ACUs to the enemy's 1 early on, and adding a chance of rushing T3 air more aggressively than before
- 5 improvements from a Radde Xander 1.0/1.0 game, mostly covering tweaks to MML and T2 arti targeting when dealing with an enemy firebase.
Acknowledgements
- Radde - providing replays forming the basis of the non-campaign related improvements, and highlighting how late M28 is to getting T3 air
-
v295 Update:
- 5 changes from Xander replays, incl fixing a bug with transports requesting engineers when idling (leading to engis being overbuilt); Raiders being less likely to run from Cybran T1 MAA; and removing a cap on MAA when a land experimental is being built
- 5 changes from Wonder replays, incl fixing a bug with land factories building engineers prior to an upgrade if another nearby land factory started a factory upgrade in the meantime; having very low health ACUs of low rated players have a chance of being sniped; and fixing a bug with T2 arti firebases not being recorded (leading to sniperbots suiciding into them)
- Mexes that aren't upgrading shouldn't be paused (as a workaround for an issue where the order to pause a unit, when given from the SIM side, means mexes generate mass for no E cost while paused)
- M28Rush should continue trying to spam t3 units after it has built a couple of experimentals.
Acknowledgements
- Radde - Various Xander and Wonder replays
- Relent0r, Ajamajan, and Grandpa Sawyer - highlighting the issue with mexes not costing E when paused and discussion on alternatives
-
v296 Update
- UEF M4 - 6 changes, including sending sweeney to the gateway and building the 3 t2 radar
- UEF M5 - 2 changes, including sending trucks to the gateway; a separate change was made to the PC AIx mod to allow 0.1-0.4 modifiers so I could check just how low a modifier M28 could use and still win
- UEF M6 - 13 changes, including building more RAS in campaign, atlantis surfacing to release stored air units, grouping amphibious units like fatboys in the same zone when preparing for an underwater attack on nearby land
- Xander replay - 8 changes, including making ASFs commit to an air fight more than before, but also being more reluctant to engage if they'd previously lost an air fight
- 2 misc changes, including that subs should surface to use their deck gun on undefended Aeon engineers
Acknowledgement
- Radde - Xander replay used for some of the above changes, and prompting me to make subs use their deck gun
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login