Turtle AI, Sorian, Teamplay - loose thoughts

[Demo]

For primarily land-based maps, turtle is not the be-all have-all solution to your problems. I noticed turtle AI being recommended a lot, because:

- it contributes to less lag (less units produced) and needless clutter (units getting stuck, bumping in each other etc)
- gets a hold on strategic missile launcher or tier3 artillery in a record time, in favourable conditions
- largely skips the phase of t1 & t2 land units spam, which the AI usually fails to make good use of (spam usually gets destroyed by other AI commander, can't breach shield generators etc), in order to rush useful units (like percivals)
- culls the diversity of t2 & t3 end-game land units to the most fail-proof, non-micro-intensive, brute-force solutions (percivals, artillery, missile launchers), making arguably best use of resources at hand
- can still turtle up reasonably well on naval-based maps and spam enemy with artillery and nukes, provided you're smart enough to disable land units production (unless you have confidence in AI's ability to transport it's land spam across the ocean) and there is enough room on the starter island to build everything needed

However, it has a couple of disadvantages:

- low priority of acquiring and maintaining mexes outside the main base past the early-game stage (air raids on enemy mexes will destroy any advantage the AI wanted to have by reducing military expenses and rushing tech3 instead), which leads to loss of map control, and in turn to loss of competitive advantage in the race towards game enders
- symbolic air & naval presence, often not enough to stop a dedicated player on either or both fronts even with reasonable resource & building time bonus (1.2-1.5)
- lowest priority of upgrading tech2 naval factories to tech3, short of rush/air AI (UEF lose arguably best conventional battleship in the game; Cybrans lose the ability for shore bombardment & submarine-immune naval base fortifications; Aeon lose their most useful ships being the advanced missile cruiser and battleship cripping their navy & bombardment ability beyond belief - Aeon AI is so bad it will send it's t2 AA missile cruisers to "melee" your fleet without destroyer/frigate support and t3 ships happen to be the only ones which are semi-useful without dedicated escort; Seraphim lose a giant punching bag called battleship, Seraphim AI was never good at making full use of it's naval units but it's the kind of blow they can't really stomach); out of all factions, only the UEF can function properly (naval & bombardment duty) without tech3, and even then they greatly benefit from it
- should it lose the main base, it still clings to "the less factories the better" template, despite it's teching ability being non-existent at that point

My favourite point of reference for AI performance is Setons Clutch, on a mirror (1 AI vs 1 AI) match, placing both AIs at the center-most spots (since it's the only spot with enough room for the mainbase without causing glitches i.e. trying to build in inaccessible location, refusing to build anything in the main base, building t3 artillery/nuker in an absurd place at the frontline etc.). Unit size set to 1000 (given AI's compulsive obessession of spamming t1 units, t1 pgens, fortifications in odd places, energy/mass storages and in general filling lower unit limits than 1000 with useless units/buildings cripping it's ability to function in late game), speed to +4 for the sake of saving time (I can have both AIs reach 1000 units and still hover at +2/+3 speed).

What I've been able to conclude testing other Sorian AI types on that map:

a) balanced AI

- this is basically a "common-sensical" turtle template, it has good naval priority without overdoing it (it knows when it's losing the naval fight), enough land units to make a punch into enemy in mid-game and enough air units to assassinate seraphim experimental bomber even when losing air superiority
- main differences to turtle is that it can recuperate quickly after losing main base thanks to the steady stream of land units and not giving up air superiority
- it's late-game logic still centers around nukes (1st) and tech3 artillery (2nd), except the transition to tech3 is smoother and more logical - it will delay transitioning to tech3 if it's about to lose it's economic ability (outer mexes, map control) - hence it will not compromise it's ability to turtle effectively by turtling pre-maturely without necessary resources to do so

b) air AI

- if left alone, it will transition to tech3 shortly after a dedicated turtle, and in late game it will only offer air superiority because there is no place for t1 bombers, t3 bombers are used in inefficient manner and t2 combined fighters/bombers primarily function as glorified t1 fighters
- it's actual game logic is early-mid game "t1 bomber rush", with a sharp decline towards turtling (tech 2 & 3 shield gens, tech 3 pgens, game enders)
- not the best turtle (doesn't tech as fast as turtle, is more vulnerable than turtle to direct land rush), not the best rusher (it puts some pressure, enough to warrant stalling balanced/rush AI in it's tech3 development, but not enough to wrest map control from them), it's generally mediocre and most-useful if you ally 2,3 or 4 AI together and this one sits in the back

c) rush AI

- with stock Sorian AI, this AI is incompetent with default AI expansion bases settings (5 land ones, 4 naval) and no longer a clear-cut winner on maps bigger than 10x10
- last to reach tech3 (it will often stall forever), last to make a naval factory (usually when it will lose it shortly after due to enemy naval presence)
- late transition to tech3 and dedicated economy-building, coupled up with highest priority not just for producing land factories, but any factories at all further minimises the air & naval presence of this AI, usually abandoning the water theather alltogether (on some maps without significance, in others critical) and losing ability to counter Seraphim experimental bombers in late game
- late teching also means this AI's main base will get pounded by tech3 artillery to the ground, or even nuked before it's ready to repel it
- it has a separate logic for tech1/tech2 land spam and combined tech2/3 late-game assault formations, because of this tech1/2 land spam will get stuck outside land factories and on the way to enemy base, infinitely waiting for more reinforcements before moving out and attacking (the way this works is that AI scouts for other AI/player's strength, once it obtains information it decides that it's able to produce X t1/2 land spam units total from Y bases - the more expansion bases it has, the more it will have to produce before moving out - in turn it will keep producing mostly-immobile troops that could be used to win the game before they're automatically destroyed due to hitting the unit limit)
- problem described above is also present with balanced AI (to a lesser extent)
- the only way to make this AI semi-useful is to be very picky where it will be spawned on the map and limit land expansions to 3, naval expansions to 2 (this will allow the rush AI to actually produce in all factories all the time and even assist them, instead of building more factories than it's able to economically support - it will also allow for a more steady stream of units made and faster transition to tech3, greatly reducing if not eliminating the "tech1/2 unit blob" problem, as well as giving it a chance at actually upgrading naval factories and maintaining a capable deterrent force); in late game it will work as glorified turtle, with sizeable "percival" spam and particular fondness for nukes/artillery

d) naval AI

- builds naval factories first and maintains enough presence to put pressure on enemy units and bases adjacent to the shore
- it's fighter screen doesn't wane when progressing to late game (which can be observed even for air AI), making this AI practically immune to all forms of airborne assault (bombers, experimental bombers, transports) and giving it ultimate air superiority
- it doesn't build many bombers/gunships, but enough to destroy isolated mexes and assist it's ground units or ships
- it builds "just enough" land units to put a deterrent in front of it's main base; on naval-only maps it's matter of preference whether to disable land units or not, it will still land transports for engineer duty or destroying enemy expansion bases/mexes
- it has arguably best priority for game enders; it builds tech3 artillery first (one of these is enough to level a base down given the nerf to shields), securing a victory instead of gambling with a nuke (which has limited usefulness and takes longer to build)
- in late game it will assault from all directions: shore bombardment, long-range artillery, nukes, air superiority, transports, ground assault
- because of ultimate air superiority and naval dominance, on land-based maps with significant water presence it's the best AI for map control, and consequently best AI to counter a turtle/air AI
- because of reduced amount of factories per land expansion base, this AI (just like turtle AI, and air AI to a lesser extent) will perform well both under standard expansion rules (5 land bases, 4 naval bases) and increased ones (e.g. 6 land bases, 6 naval bases)
- usefulness of increasing a naval base limit is debatable; this AI will often assist it's naval factories, especially on naval maps, whereas over-exposed and poorly-defended outer naval bases would only be a waste of resources and time for AI's engineers (similarly, it's nigh-impregnable fighter screen makes number of land bases largely irrelevant on naval-only maps)
- not suprisingly, it's the best AI for naval maps; however, it will still function poorly at shore bombardment if AI markers are nonsensical and won't develop properly in late game if there is not enough room at starter island (more like general AI limitations than faults of naval AI itself)

----------

"The legendary unit blob"

Curious whether the "unit blob" problem crippling rush AI could be fixed, I decided to install Duncane's AI fixes to SupComFA's data folder, seeing as Sorian himself once recommended installling it alongside his own AI (playing Sorian AI profiles influenced by Duncane's code). The result was that indeed, unit blobs were gone, and for good. Suddenly rush AI on Setons pummeled through air AI, naval AI, balanced AI and turtle AI's main base with brute force before the late game truly began, and before the game ended it had a tech3 artillery and a nuke ready to pummel through the remains of what was left. As a bonus, properly-working rush AI is arguably the best AI at map control & economic starvation on land and primarily-land maps, covering the whole map with tank columns destroying any and all attempts to capture mexes freely otherwise.

Unfortunately, this little experiment left more questions than answers.

Would the host of a teamplay game desync everyone not having Duncane's files in his installation (only the core file was needed - 00AIFix.scd)? I haven't tested it myself yet, and it's definitely curious whether host alone is responsible for AI simulation, or would lack of these files in the client computers cause problems as well.

Do the Duncane's fixes, besides fixing the unit blob problem, affect the Sorian AI in unintended and detrimental way, in terms of it's actual fighting performance? That would require a thorough test, I'm not ready to commit myself to.

Would Duncane's fixes cause problems with Nomads or other mods? It could turn out, they're only compatible with stock 3599/3603 FA and FAF, as opposed to mods adding additional units or factions not necessarily covered in AI's logic.

EDIT1: Turns out above-mentioned files won't cause a desync if there are no AIs in the game. If there are, all players must have the correct file sitting in the data folder.

Finally, the "fix" itself is not an optional plug-in like GAZ_UI, but goes directly to the data folder, in turn upsetting the core game and not just FAF. Therefore, is it even worth consideration for a future FAF patch? It does fix a whole AI profile under FAF, and AI profiles/teamplay maps are arguably one of the FAF's features, but definitely not the developer focus. Implementing something unverified, with unknown consequences and that doesn't work unless placed in data folder rather speaks against it; which is a shame, because the unit blob problem is indeed holding back the rush AI from being competent, and giving teamplay players a new toy to play with (I played with unlimited AI expansions to verify that it does address the core problem rather than it's symptoms).

Tired of turtle AI? You're in bad luck, unless you go rogue and install Duncane's files on your own risk. What you can do is put that small file in ~Supreme Commander - Forged Alliance/gamedata folder, host a passworded teamplay game, get verified people to join it and test whether it's worth your effort. When you want to play something else and not cause desyncs, all you have to do is remove that file after exiting the teamplay game you hosted; lobby itself won't care that you copied or deleted a file. Lazy way: make a "backup" folder in ~/gamedata folder, keep your AI file there when you're not hosting teamplay with it, swap the file back to ~/gamedata when hosting again.

Point of reference: http://forums.gaspowered.com/viewtopic. ... 76023be0e9

And specifically - http://members.iinet.net.au/~dionysus/misc/AIFix.zip (00_AIFix.scd - the only file you need)

[Astrofoo]

I can confirm that I have Duncane's AI fix installed and I do not play teamplay AI games at all. I play exclusively multiplayer. It does not cause desyncs. I have played about 120 online multiplayer games without desyncing. Feel free to install Duncane's AI fix without worry.

However, if there are AI's in the game it will cause desyncs if the other players do not have the AI fix.

[Demo]

That's a relief, thank you.

Now that one problem is sorted, I would be more than elated to hear if there are confirmed, wider repercussions in terms of AI performance save for the AI blob problem (which in Duncane is fixed by sending idle/stuck units at an enemy in loose formation, instead of them trying to continuously reform & wait for reinforcements - tech 2/3 "formation" troops still function normally, without delays, in cohesive groups).

[eXcalibur]

why not include this in the upcoming patch 3622?

if it causes more problems than does good, the changes can get reverted. that is almost the same method as with most balance changes...

[AwarE]

I find it strange that you use Seton's Clutch as an AI testing ground. I would think that this map has the worst full set of FA AI markers compared to all other. The land movement markers tell the AI to path straight into a cliff in both north and south. Same goes for amphibious markers, although these have many other problems where they meet water.
Really m8 it has a horrible set of FA markers and I could go on and on pointing out 100 things that can be done better.
Just because the Aeon Sorian AIx Turtle runs badly on Seton's means nothing. Put that excellent AI on a map with good markers and with multipliers @1.2+ , then you will be amazed at its performance against any other AI accept the same AI [Sorian Turtle AIx] only different faction ... UEF.

Its funny, I do agree with most of what you say ... but your wrong if you think any other AI can beat a Sorian Turtle AIx on a map with a correct set of FA markers. I can say this because I have tested over 5,000 AIx games. I build the maps for the AI, I set the paths for the AI, I add restrictions to give the AI its best chance to kill at least one human commander in a TeamPlayAI match.

Unfortunately big naval battles require higher multiplier settings. I aim at 1.2 as the normal for naval maps. Better players would need a higher multiplier or they would easily crush the AIx water build.

As you point out most other AI build too may T1 units, even when there is no land link between bases. In the end game the Turtle rules because it has the best economy and builds a range of game enders like nukes, LRCs and experimentals. You know you had the multipliers set at the right level to match your skill if you killed more than 50 GCs and had 200 structural losses. Most people make the mistake of adding too many AI instead of just two, that are teamed and placed in the correct start locations. Sorian Turtles AIs need a position to expand to or they will not run well.

roj
PS: I don't see the problem that Duncane is meant to fix.

[Demo]

Resource multipliers and purely-land or mostly-land maps will always skew the game towards turtle, by reducing importance of land-grabbing and extending the travel distance between opponents, making t1 and t2 land useless against enemy base. Rush AI (when working properly), assuming a land passage to enemy base, is about the only AI able to crush the other AI or player before mid-game, with or without resource bonus depending on the player's skill.

My beef with turtle is it's over-emphasis of economy over map control early in the game, combined with over-emphasis of exponential base growth in weird directions basically following a pattern of "t3 pgen - t3 fabricator - random experimental or two - repeat). With Rush AIs and their variant the same pattern is slowed down by considerable factory output, which seems most sensible in the case of land AI due to the sheer mass of numbers pretty much neutralising any lone experimental sent by the turtle at that point.

That may be, inability to properly use experimentals (or even build some, like aeon t4 battleship) by combining them with own force seems to be a culprit here. Your GC example is basically a variation of land units spam replaced by insane amount of GCs in cheated economy - GC is about the most idiot-proof, AI-friendly, effect-to-cost unit in the game (spam of monkeylords mass-for-mass would be ten times less effective even if the AI's tactic of using them would be the same).

There was a brief time when I dabbled in modding another game on AI/gameplay project, after many revisions talks have come to a point that with the stock parameters and limitations of hardcoded game mechanics, the AI core we tried to influence has pretty much achieved it's peak performance with the base game, and it was far from satisfying. What I proposed was to start designing the game around AI's limitations instead of designing the AI around current game system, and then create a set of guidelines for other modders to follow on how to design their mods with our AI in mind. The result was a superior experience in terms of challenge, but also a reduction in strategic and tactical choices for the player.

I believe in your effort of brainstorming effectiveness of various AI profiles and designing various teamplay maps you've settled to balance maps themselves around turtle AI performance, given that it was most fail-proof and had no less than 2 factions excelling at it (namely UEF/Aeon for most of the time - I believe Seraphim is worth consideration as well). Similarly, aware of player being able to generally out-smart and out-play the AI the longer the game goes, I'm not surprised you recommend giving the AI at least a mild economic boost and placing it carefully on the map so it builds the economy and functions properly for as long as possible, before it's ineffective economic line-up is no longer able to put up a serious challenge. Surviving a strike of 50 Galactic Colossi surely gives an impression of Tower Defence games I've had over battle.net's warcraft III custom maps, and despite their inherent simplicity of design these were among the most challenging types of gameplay available.

The difficulty here is drawing a line between improving/supporting the core game (whilst aware of it's limitations) while preserving original character of the game (multi-front, balanced, all-tiers-viable warfare with air, sea, water) and actually designing for highest challenge irrespective of the variety of the gameplay provided and original game flavour. These are two schools of design following different philosophy - first assumes the player will employ "house rules" to give him a balance of freedom and challenge, the other that player will be ruthless and abuse every AI handicap or loophole in order to win and thus the system itself needs to limit what the player can reasonably down to the level of AI in a mostly pre-determined setting (guaranteed challenge regardless of house rules). In your example, tinkering with map design, setting cheat/omni multipliers, turtle AI with XYZ factions and ABC rules of AI placement on the map heavily narrow down effective strategies for the player in terms of beating the AI, making a more streamlined experience regardless of player expectations past launching the match with pre-defined settings.

PS: I don't see the problem that Duncane is meant to fix.

Here is a Rush AI vs Rush AI battle, the dreaded Setons, pure Sorian AI. 1000 units limit, unlimited firebases.

AI blobs: units that get produced in the t1/t2 disorganised land spam, then assembled into bigger forces before being sent to the enemy. Size of the group before it's ready to move out is directly proportional to the number of available fire bases, potential fire bases (total limit) and calculated relative own strength to that of the enemy. The more firebases, the thicker enemy defence, the higher enemy tech level and the more even the fight between two opponents, the more likely is the AI to assemble it's "AI blobs" endlessly, effectively freezing them in place instead of hurling at the enemy. In most cases, this loses Rush AI the game vs Turtle AI.

AI battlegroups: very rare, combined t1/t2 formations and all combined t2/t3 formations, that get assembled shortly after leaving factory and then head across the land markers towards enemy base without failure.

What we have here:



A blue blob, one of many, this one sits in the base and does nothing, while mexes in the center of the map have been destroyed aeons ago and enemy is pushing towards main base. AI in the bottom over pure Sorian AI has much less inclination for attack than the top AI for some reason as well, presumably because top AI pushed with it's commander inflicting casualties, increasing the "sufficient strength of blobs before sending them out" threshold for blue AI blobs despite inferior economy, making the blue AI stall blobs even longer than red AI before commencing them to fight, in turn losing game as a result.



Red blobs. Nothing more to say, they sit idly by before their t1 counterparts are self-destructed due to hitting 1000 units limit. They're mostly idle till the decisive battle for blue's main base is mostly over and red's economy has advanced well beyond t3, making organised, not-blobby, combined t2/t3 stacks only.



An example of successful, combined battlegroup of Sorian AI. Most common sight in late game when only organised, combined t2/t3 groups are produced, and "stalling blobs" problem is generally over.

Another battle, turtle AI on bottom rush AI on top. Same map same settings, except this time with Duncane.



A steady stream of rush AI's units. There are small blobs, which unlike pure Sorian AI have internal "timeout/cooldown", which hurls them to the enemy base in loose formation once the timer for "assembling the blob and making it bigger" expires. Essentially, a fail-safe trigger which forces the AI to throw it's idle units at the enemy if they've been idle for too long. This feature is completely absent in stock Sorian AI.



Successful mid-game blob of t1/t2 units at 25th minute mark roflstomping turtle's base. Also, an almost-dead blue turtle commander. Nuff said.



Nighty night, turtle commander. Rush:Turtle, 1:0.

PS. I've also tested games over Duncane-influenced Sorian AI profile, setting rush AI at bottom and turtle at top. Same result, turtle got baked. Markers are not the problem here, they might be idiotic but they don't favour either side of the map in terms of pathfinding. Ineffectiveness of Rush AI is directly related to the "endless blobs" issue.

PS2. Also compare, Rush AI penetrating turtle at 25th minute over Duncane versus Rush AI about to attack main base of "hopelessly blob-stalled" Rush AI at 48th minute over stock Sorian AI. Basically, Rush AI's effectiveness at least doubles just because blobs are gone. On a side note, Rush AI vs Rush AI fights with Duncane tweak are interesting, because they're like tug-o-war between two forces until little, incremental advantages eventually turn the tide of battle.