SMACX Thinker Mod

[bvanevery]

I suppose that part of the formula works as intended, e.g. limits certain exploitative strategies.

Until you get everyone to legalize chemical weapons.     Which can be surprisingly easy or surprisingly hard to do.  Depending on what, I don't even know.  Dumb AI thoughts about weapons it's not even gonna use?  Anyways, gotta love the morality of Planet giving its blessing according to the United Nations protocols.

It doesn't make sense to say eco damage gives some huge advantage to the AIs since human player can easily exploit it.

How much Super Hover Former empire repair have you really done?  It gets old.  Especially those flooded river plains that actually kill any unit you leave on them.  I don't care how much benefit you're supposed to be getting by fighting the good fight against mindworms, it's really a drag to have your empire ripped apart again and again, never being able to make any progress on it.

And that's assuming you still have land to work with.  Once the world turns to water, your minerals drop like a rock!  Even the Pirates can be affected, because deep ocean doesn't generate any minerals.  Once you have no productivity, you don't have a way to get out of your situation.  You die.

I learned the hard way that I actually have to premeditate about that.  The climate scientists of Planet have to be really really good.  They have to anticipate that they're going to lose all their minerals and start picking city sites and production strategies early to counteract that.  Like, if you built a lot of mining platforms and subsea trunklines in advance, you'll be ok.  If you didn't, well screw you, you're just not a good scientist are you!

You also have to premeditate about how to survive the mindworm attacks.  Like build the Neural Amplifier before you set off the world destruction.  There's a lot of stuff you have to do, it's no free lunch.  You also spend all your time fighting mindworms and none of it fighting other factions.  That can turn into a real problem.

Never once, to this day, have I survived a chemical weapons triggered mindworm onslaught.  And the only reason I survived Fission Armor's early factory debacle, is because Tim made the Ascent to Transcendence take super duper long to produce.  If it had been the stock game's version, or my version, the game would have been over with complete AI victory.  When I played my 1st Thinker mod game combined with my mod, I knew that endgame was coming, which is why I quit.

What maybe could be done here, is have some kind of a config option to have the AI adjust the extent it uses heavy terraforming improvements, like boreholes and condensers. It could even adjust it according to the faction definition, but I'm not sure yet how that might be implemented.

What I actually want, is for the AI to do what it's already doing, and suffer for it.  The same pain that I get.  Because then we get to see, this isn't some kind of best / obvious / uber strategy for the AI to be doing.  The AI could easily make all these great productive minerals and then burn out.  Or maybe it will actually build some eco-friendly facilities, I've actually seen the stock AI do that, when testing the early factories debacle of Fission Armor mod. 

But the point is, the AI will not simply get to cakewalk human players, like we're all some kind of pile of dummies who don't actually know how to play the game, and should have been using the same strategies the AI is using.  The game is set up to cheat like a mad dog for the AI, the AI doesn't pay the cost of its ecological transgression at all.  That was fine when the AI wasn't any good at supply crawlers and boreholes, but you've totally changed that.  You've given it the knowledge of a power gaming minimaxer.  Well, such a "player" needs to suffer the consequences fairly.

I'm not interested in my life being easier with the eco-damage, like dropping down to Librarian.  I think the eco-damage, as is for humans on Transcend, are ok rules.  I've been playing with those rules forever.  Dropping to Librarian is basically nerfing eco-damage rules out of the game.  So now you've got your AI, and it's doing its horrible crawlers and borehole thing, forcing me to play that way to keep up.  Because it's now the golden path through the game.  I don't want there to be only 1 golden path through the game.  I don't want to be forced to play the game the AI's way.

So in my perfect world, the AI has to actually play the game that I'm playing, and not be given a huge eco-damage cheat.

[bvanevery]

and one of the terms in the final eco damage number is "3 - PLANET", and not "PLANET + 3".

I fixed it.

[Induktio]

Just for fun, here's a talk called Playing to Lose: AI and "Civilization" (Soren Johnson). He was the main AI designer at Firaxis from Civ 3 onwards.

Btw, one of the lines on that revised eco damage should be like this:

Improvements = ((2*WorkedImprovements + 1*UnworkedImprovement + 8*Boreholes + 6*EchelonMirrros + 4*Condensors + Seabase - Forests) / (8 * (1 + TreeFarm) * (1 - HybridForest))

A very important detail in the ordering of the brackets.

[bvanevery]

Btw, one of the lines on that revised eco damage should be like this:

You mean on one of the wiki pages?  You could just edit the wiki to be correct.  It doesn't require any kind of special permissions to do that, just an account here.

I skipped through the Soren Johnson video.  I didn't really get a lock on anything I'd especially comment on.  In general, the problem of an AI being written that simply kicks the player's ass, and thereby isn't any fun to play against, is a known game design problem.

[dino]

In general, the problem of an AI being written that simply kicks the player's ass, and thereby isn't any fun to play against, is a known game design problem.

Giving lame AI so many cheats, that it can crush a player mercilessly no matter what is indeed unfun, unlike competent ai with little cheats.
So I call it bullsh*t, if we are talking actual good AI, not AI cheats, since I've never played a 4x game where too good AI was a problem, opposite was always a prablem,

Most probably 90% of players never spend enough time with 4x games to learn to play competently enough, to require competent AI to enjoy the game, so producers can't financially justify spending more resources on AI development and what we were getting was considered good enough for overwhelming majority of gamers.

Other less likely explanation is it's actually so hard to do, that no one ever succeeded despite trying and allocating serious resources to AI. This one may, or may not be true, we will never know, because probably spending something like 25% of a game budget on competent AI programmers in order to make really great one, was never even attempted by any game studio, since it wouldn't be considered a smart bussiness decision.
Imo it was always a financial decision, not a game design decision, but a small vocal minority of hardcore fans need to be addressed in an interview, so we get "bad ai is actually a good game design" argument...

Short version: Average gamer is too dumb, or lazy to care enough about AI, so a good one would brought far less revenue, than it would cost to develop.

[bvanevery]

What I said is nevertheless a known concept in the game industry, and is not limited to 4X games.

Budgets have been a historical issue, as have computing resources.  I don't know whether either are issues lately.  I haven't been keeping up, as it mostly doesn't matter to me what the rest of the industry is doing with AI.

The only data point I've collected in that regard, is Gal Civ III is reputed to have tried to really sink some 64-bit threading resources into AI.  The player feedback I read in /r/4Xgaming, doesn't seem to say that yep, they obviously succeeded.  I haven't been willing to embark upon GC3 learning and mastery to find out one way or the other.  Got other development priorities right now.

Multiplayer bias is an ongoing issue.  The idea that since humans are "always" going to do better than AIs, that human players would prefer to seek other humans to play against.  This claim leaves out a substantial chunk of players, like myself, who do not enjoy interacting with other humans online.  Either because humans turn out to be too much into hardcore minimaxing to be any fun to play against (a subset of the general lobbying problem), or they're just jerks, making Life more unpleasant than it needs to be.

Multiplayer bias is further encouraged by social media marketing models.  Which again, I personally don't like, and do my best to avoid participating in.  I'm a Facebook refusenik, for instance.

Aside from those problems, is game production bias, which Soren Johnson did comment on at one point in his video.  The game designers are out in front with all these rules and systems they keep heaping onto a game.  Human playtesters egg it on: yeah yeah yeah MOAR MOAR MOAR!  The AI programmer plays catch-up to those concerns, in an industrial studio setting at least.

As a solo indie developer, I have the power to say "no" to that.  To recognize that half of what's in SMAC, is game mechanically useless.  I've never seen anyone endeavor to put a 4X franchise on a diet.  I've only seen developers follow the mantra of "MOAR".  I intend to do differently.  My SMACX AI Growth mod is like a dress rehearsal for what does and doesn't matter about the game.

[Induktio]

Upgrading an engine to 64 bit is really only significant if one has to allocate more than 2 GB memory. That's what the purpose is. Rarely do these AI routines perform such heavy computation that 64 bit would help there. Most of the time we're mainly dealing with scripted routines, not some computationally heavy machine learning stuff.

Usually something like pathfinding might be a bottle neck, probably also in SMAC it's the culprit that hangs the game from time to time. Though that might have to be investigated further with a profiler. Interestingly I created a separate, simpler pathfinding routine for Thinker which seems to perform its role very efficiently. From very early on, I figured it was necessary so the AI would have an idea how to deal with the complex and modifiable terrain.

[bvanevery]

Most of the time we're mainly dealing with scripted routines, not some computationally heavy machine learning stuff.

GC3 may have attempted to do some computationally heavy, threaded, machine learning stuff.

Interestingly I created a separate, simpler pathfinding routine for Thinker which seems to perform its role very efficiently.

The main thing I hate about the stock binary's pathfinding, is the way it makes mistakes with Formers and other non-combat units.  Get near a border, and the pathfinding goes crazy!  If it would just use whatever the combat units are using, all would be well.  Then you just get the occasional total meltdown where a unit moves back and forth until it asks you if you want to quit doing that.

[Induktio]

> GC3 may have attempted to do some computationally heavy, threaded, machine learning stuff.

It would be good to have a source for this kind of claim. If they were using something like that, I'd guess it's a pre-computed model. Large strategy games don't necessarily lend themselves easily to that kind of analysis. The range of possible actions is very large in those games, compounding the problem.

About the other issues raised in this thread, I won't give any promises what stuff will be in the next version of Thinker. Mostly it should deal with fine-tuning the existing behaviour though.

[MercantileInterest]

What does Thinker do to the faction placement algorithm?

[bvanevery]

> GC3 may have attempted to do some computationally heavy, threaded, machine learning stuff.

It would be good to have a source for this kind of claim.

Search /r/4Xgaming and also /u/draginol.  That's Brad Wardell's account.  I remember reading something about this, but not caring enough to get into the details.  First I'd have to see if it plays well, and I was unwilling to put in the time.  Only if it was beating me, or challenging me, would I bother to figure out how it was doing that.  So I read what other people said about it, and the results were mixed.  Like, some people said it wasn't particularly smart and definitely relied on resource buffs to provide challenge.  I don't actually know one way or the other.

My memory of 2 years of posts in /r/4X gaming is also, that nobody else has done 4X AI worth a spit.  GC3 is the only one that is worth investigating for its technology.

Mostly it should deal with fine-tuning the existing behaviour though.

You really don't need roads on every single square.  It's a waste of terraforming production, and it makes AI territory easier to invade.

You don't need a Condenser like every other freakin' square.  Personally I have preferred to put them on Rolling, Arid squares.  Then I get a mineral to go with my food and complete lack of energy.  I never put mine on Flat terrain because then you just get a pile of food with no minerals and no energy.  If squares are already Rainy, then they don't need Condensers next to them.

[Induktio]

What does Thinker do to the faction placement algorithm?

Well, it is described in the documentation, the new algorithm decides where all the factions will spawn on the map. The main purpose for this is to prevent multiple factions being bunched up in a small island or similar. That part the algorithm does really reliably, it will evaluate the size of the starting landmass and discard bad locations. Random map generation also works better when it's paired with my WorldBuilder enhancements that are now also available from the project page.

And yeah, I looked at those testing reports. It might be getting redundant to test the same parts of the game any further though. But then again, the next version is not ready.

[dino]

Compared to vanilla, builder gameplay is nailed, but thinker has now opposite problem to vanilla it's extremely passive with no defences.
Enemy territory full of boreholes and condensers, tightly packed with bases and roads is like free candy and leading faction can be easily overrun in few turns in late 2100s.
It's literally easier to conquer the world than vanilla and I saw bvanevery here and one guy on rpgcodex having exactly the same experience with thinker.

My thinking is that you can't make ai capable of determining when it can safely buildup and when it needs defenses, so it has to do both all the time and rely on cheat bonuses to compete with more adaptive human.

1) I've proposed already a formula that it needs at all times to have ( 50% + ( ai_fight * 10% ) + 10% if vendetta + 10% if enemy units inside borders ) of mineral production used by unit support.
2) It needs a way to keep the units upgraded, from my testing vanilla ai do upgrade units when it has excessive amount of money, so maybe some change to make it more of a priority and reduced upgrade cost for the ai.
3) It shouldn't build so many sentinel type of units.
4) If player military power variable ( used for military ranking ) is 3x smaller than ai and ai mood toward player is bad enough it should just trigger vendetta straight away.
    Under some conditions vanilla ai can just declare vendetta without going into dialogue ( meybe it's in perihellion only ), so it's an idea to replicate this when these conditions are met.

Whathever you have on your plate currently, you should put it on hold and try to address issues above first, because currently it's boring and passive in early game and easy to conquer.

Another issue is ics style of bases placement, it doesn't compete for land in early game I can easily and without conflict, grab more land early, than against vanilla ai with exploration focus.
1) Maybe you could incorporate vanilla colony pod ai into yours, like give it 40% chance it will behave like vanilla, 60% chance it'll use your movement/ base placement code.
2) Make it spread bases 3 tiles away, never 2 I can give ai cheats that makes it popboom, which will be wasted with too tight placement, also tightly packed early bases have less special resource, monoliths, rainy tiles to work with, plus the issue of not grabbing land fast enough early.

As amazing, as builder aspect of the ai is in thinker, it's as unsatisfactory to play against as vanilla currently, just for different reasons.
It can compete only if you don't rush them. If you don't attack it, you can either give it enough cheats, that you don't have a chance, or not and you'll always win a very predictable and uninterrupted builder game.

[MercantileInterest]

A pair of variables from Yitzi alphax:


Disable, ; Technology to convoy all three resources
Disable, ; Technology to convoy all three resources outside of base radius

Convoying all three resources doesn't sound like a useful gameplay feature but this would be:


Disable, ; Technology to convoy resources outside of base radius

With this option, crawlers could be a very early game tech but not overwhelmingly powerful.

Another Yitzi variable:


0        ; If non-zero, borehole count per base is squared when determining ecodamage {under -2 billion to over 2 billion}

Makes it eco-suicide to spam more than one borehole per base. This would interfere with Thinker's current functioning.

Right before he left, Yitzi talked about the need for huge armies of formers bogging down gameplay. His proposed solution was to make all formers cost mineral an extra mineral upkeep regardless of clean reactors or support value but at the same time to reduce the time needed for terraforming in the alphax file.

[Induktio]

Umm yeah, recently progress with the mod has been somewhat slower than I would have liked. I've evaluated several new changes to the terraforming algorithm and scrapped some ideas along the way. It's pretty difficult to code robust heuristics that work well in a wide range of situations.

> Compared to vanilla, builder gameplay is nailed, but thinker has now opposite problem to vanilla it's extremely passive with no defences.
> 1) I've proposed already a formula that it needs at all times to have ( 50% + ( ai_fight * 10% ) + 10% if vendetta + 10% if enemy units inside borders ) of mineral production used by unit support.

I've been wondering about this as well, it looks like currently Thinker builds too few military units. Luckily this can be modified easily by adjusting the threat formula. Currently the main heuristic it uses is the distance to the nearest enemy base. The formula has a wide range of variables but I think it does make sense to add "if enemy units inside borders" factor there. We'll see what kind of effects this stuff will have when I get the next develop version ready.

> 2) It needs a way to keep the units upgraded, from my testing vanilla ai do upgrade units when it has excessive amount of money, so maybe some change to make it more of a priority and reduced upgrade cost for the ai.

I don't remember seeing the AI upgrade any of its units. Can you provide some kind of a save game where this happens consistently?

> 3) It shouldn't build so many sentinel type of units.

It's just currently it builds too few units. When its adjusted to build more, the units should be mostly offensive in nature.

> 4) If player military power variable ( used for military ranking ) is 3x smaller than ai and ai mood toward player is bad enough it should just trigger vendetta straight away.

Diplomacy stuff probably will not be a part of the next update. It will need a lot more reversing before anything can be implemented.

> 1) Maybe you could incorporate vanilla colony pod ai into yours, like give it 40% chance it will behave like vanilla, 60% chance it'll use your movement/ base placement code.

This colony pod argument is something I will mostly disagree with. Grabbing territory doesn't do much by itself unless it's populated by bases. Closely packed based can be defended more easily anyway. Each turn a colony pod spends on movement is a turn of missed productivity from a future base.

> Disable, ; Technology to convoy all three resources
> Disable, ; Technology to convoy all three resources outside of base radius

Would make crawlers super-overpowered and require lots of patching. Nah.

> Disable, ; Technology to convoy resources outside of base radius

Insignificant, Thinker mostly crawls tiles that are inside base radius anyway. Thinker formers wont even consider improving tiles that are not workable by some base.

> 0        ; If non-zero, borehole count per base is squared when determining ecodamage {under -2 billion to over 2 billion}

Not planning any patches on the eco damage formula currently.