I looked at an AI in game, I read what you guys wrote here and I see that AI is mainly focused on setting correct hendicap. This is wrong direction.

I see following main issue with the AI:

1. Bad planet management focused mainly on population growth (believe that program is calculating efficient production, research and weath and with current bonus from hendicap it gives him best results), AI is not focusing planets, in result in late game i have at least 5 Times higher research prod and wealth than computer, writing function for focusing (how offen, which planet and when) is preaty easy so i don't understand why this is not happening. AI is not using bonusses from adjected tiles, also this should be easy to fix. And lastly does not suport its planets with economy starbases, however always creating shipyard even when unnessesery.

2. Fleet management is awful. hundreads of ships killing it self one by one on my few ships. No strategic movement to invade my words from behind. Ai singe ships that coming to close to my much stroneger fleet just to attack basic constructor, or with fleets that  are not fully using its logistic sizes. (Ai is producing a lot of ships mainly cheap and weak and later AI has no money to upgrade it, here i see place for handicap to allow AI upgrade its ships much cheaper than for human player), also becouse of poor planet management their research posibilities are not grothing as fast as new research cost and they stay with the same ships for most of the game. Moreover they produce lots of transport ships that slowing down their production (population taken to ships) and just standing with them or sending them directly to my fleet to be destroyed.

 3. Current hendicap is creating bad playing expirience. Very strong AI on begin, (that you can anyway easly resist thanks to prototype weaponry) and later on very weak. Moreover you can easly defeat even hardest AI by reducing numer of planets in the galaxy. Computer difficulty is directly related to the number of planets. I even played with race editor and checked that the best computer race is the one that has bonuses to population, farming and morale, focused on exploration/expansion and military.

I'm missing a bit of fineness in AI. This game should be a bit more than just "lets colonize as many as we can and do a brainless zerg" 

double post

I don't think that's true. I know Frogboy wants the best possible AI for the game, he spends a lot of his time working on it, but he's primarily the CEO of a company and he has to invest the available resources wisely. Giving us a god-tier Deep-Blue type AI will not provide much return on investment, and while that's a shame, we (the fans) actually do have an opportunity to step up and provide some support on that front.

We're helping out here for fun, because we want to improve the game, and want to improve our and our fellow players' enjoyment. If anyone followed Supreme Commander (almost a decade ago now) you may have heard of the Sorian AI. Sorian was a modder who totally overhauled the AI system for that game (and its expansion, Forged Alliance), vastly improving it without handicaps. Eventually he was even hired by GPG and then Uber to work on AI professionally... But his mega-AI was only really for hard-core players. I expect most of the SupCom playerbase never felt they needed awesome AIs.

If we here have the drive, we can develop the capacity to upgrade the AI and improve all our experiences.

Over the past few weeksI've learned to read the .XMLs, so now I can squash issues where the AI is un-intentionally handicapped by code errors. It's interesting, and fun to learn about coding. Next maybe I'll go through GovernorDefs.xml and take a whack at how the AI builds its planets. 

I invite you all to do the same, that part of the coding really is quite straight-forward. Look in AIStrategyDefs.xml and see how the AI prioritises fleet builds, look in TreatyDefs.xml and see how the AI values various diplomatic options, look in GalCiv3AIDefs.xml and learn how the AI determines at what range it considers opposing ships to be threatening etc etc.

As Frogboy said: 

Let's see if we can trace the code well enough to see why the AI is making bad fleets, or prioritizing research in a sub-optimal way for its current goals. If we lighten the load on those kind of fairly easily accessible annoyances, it frees up Dev time to take on the things we can't change so easily by altering some variables, like AI governors not having the capacity to specialise planets and the other stuff that we all want.

To build on TurielD's fantastic post. 

I've always assumed that the AI ship designer had a plan or a goal when it designs it's ships and because it's an AI wouldn't have any trouble keeping up with which design is for which purpose.  Presumably the AI API has the ability to assess threat types and ship ranges.  I woudn't have thought that it always "fills up" any available space on ships especially transports and constructors just for the sake of filling space and if it did, then do as I feel like most humans do and fill it with engines. 

For example, AI "decides" to take PlanetX. Dispensing with any FOW issues, it assesses PlanetX's defenses; how many ships in orbit or in range of X, how many people on X, what's my invasion level vs the target, even using invasion success routines to predict it's chances.  It might shoot for a 75% or better chance of successfully invading using a combination of invasion techs and sheer numbers to achieve that probability. 

That then leads to it's invasion plan. 

1. Get space superiority. Design and build appropriate ships with the right weapons, defenses, range and speed to control the area around the planet. Do I have what I need? If not, then build it. Keep control for a couple turns before letting the invasion fleet get too close. 

2. Design and build the appropriate number of invasion ships (and some escorts) and have them move into range of the target. Range = 1 turn of movement. 

3. WIN. 

Now, I'm assuming that the AI and more importantly SD / Brad already know this and there are all kinds of holes and gotcha's from this simple approach that I'm not thinking of. 

However, if the primary problem is as TurielD's analysis shows, then perhaps these things can be fixed by helping the AI design better ships and more importantly, as the database of player designed ships is assimilated, helping it CHOOSE among the many competing designs. 

None of this post (as opposed to TurielD's) does what Brad asked for; specific examples of bad behavior. Hopefully it will be helpful for conversation purposes. 

Regards. 

AI is one of those things where it is very easy to know what you are talking about, very complicated to define and almost impossible to achieve. Only women, I believe, are more complicatedly opaque than the field of AI.

One of my old favorites (as I've said) was the SSG series of games. Part of their approach to making a good AI - and they made some really tough opponents - was to limit the player to making high-level decisions and let the AI make the lower level ones for itself and for the human player. This limited the options the AI had to contend with. For example, in 'Battlefront' (if I recall correctly) a division was split up into regiments or brigades, with up to 3 battalions that could be cross-attached (like divisional artillery). Instead of maneuvering each battalion separately the player set general orders for the regiments and the AI maneuvered the battalions. Sometimes brilliantly, and sometimes...

Now, for people who love micro-management this was a much-hated Hell, but the game played superbly, delivered reasonable results and gave you a lot of sympathy for senior officers in their attempts to get their juniors to do what any reasonable PERSON...

Ahem. As I say, it produced results I still remember today - I'd still count it as a great game and a great AI - and it made you really - really - think about your plans and about how easily they could go wrong. (My sister described being a successful restauranteur as 'graceful recovery from imminent catastrophe' and I think that applies to generalship as well).

For my part, I don't want an AI that plays the most efficient game possible. I want something with personality, something that fights a war in an Altarian way, or in a Yor way. I want the game to be memorable because of what the AI does, and what I do to counter, and because of what IT does to counter me. After all, no nation in the history of the world ever fought a war of perfect efficiency. Nations (read people) always have ulterior motives and hidden agendas, prejudices and tendencies. I do want an AI that is no pushover, but mostly I want something with its own peculiar personality - something that plays to the Turing-test side rather than to the efficient analysis side.

Unfortunate AI behaviour: Sending damaged ships out on their own as vanguard of an attack force.

This relates to my previous post, the AI doesn't seem to want to fleet up; it's so averse to it, that rather than join the ship 7 hexes behind it, this ship with 1/3 HP is rushing headlong toward an enemy world. 

I think it would be good if the AI considered its own health. For instance, if health < 50%, it should seek out the nearest (military) fleet with enough logistics capacity to house it. This would help the AI a lot with combining forces I think, and prevent the death-by-a-1000-cuts thing that can happen to larger vessels. Perhaps if health < 25%, it should attempt to reach the nearest shipyard to repair, and/or wait there for the next military ship to come off the production line and join that.

To an extent, yes. It really depends on what you want the AI to do differently - kinda like this XKCD strip: http://xkcd.com/1425/ 

It's easy to change what ships the AI builds, what improvements it builds on its colonies, how it prioritizes research under certain circumstances, how much it cares about other factions' ships getting nearby etc. But I haven't the first clue about how to tell the AI it should focus its vessels on taking a single world, or only making starbases inside its own cultural influence zone.

It may be that those are modable, and I just haven't found the right file and code line yet... or they may take a full rewrite of the AI from the bottom up.

Quoting TurielD,

reply 110

Unfortunate AI behaviour: picking rally points in enemy territory, rather than gathering ships and attacking together.

 

An example from a soak:

 

Reduced 29%

Original 1920 x 1080

 

 

Running all AI on normal difficulty here, my Doofus AI faction is at war with the Krynn. As the Doofuses are adaptable and expansionist, they're about equal in power and production, so we can reasonably expect both sides to field roughly equal fleets. 

 

What I'm seeing is now that the AI chooses to direct ships towards the enemy (Bunch III, down the bottom) it's sending individual ships to meet at a rally-point which is actually past the enemy world, from where either ship has started out. This means they will easily be picked off 1 by 1 if the Krynn have a 2-ship fleet on the defense.

 

Save Game:

https://dl.dropboxusercontent.com/u/45479330/Doof%20Soak%20Bad%20Rally%20Point%20Choice.GC3Sav 

Awesome, you put proof positive to what I said early on rally point during a soak. Unfortunately, I wasn't thinking about screenshots and saves. So all kudos to you sir for proof!

<Edit> that is why you should never ask most programmers to do QA. I see things at work and go to fix them, but here I usually cannot. As to my own work, I don't WANT to hear it breaks something!  </edit>

think it would be nearly impossible for an AI to be perfect the game has too many possibilities of what to do. but the AI should not cooperate when being exploited by the player. surely this is or can be frustrating for new players when they don't get what they're asking for, but such things could be part of higher routines that aren't used for beginner, normal diff.

another prob is with engines and range. the AI put so much range on ships sometimes they fly in regions completely irrelevant to realistic strategical goals. it would be better or safer for them to be more territorial. engines often loose their bonus if the ships are fleeted together esp. with tiny small craft. the AI needs a routine to examine it's own fleets, making decisions to build variable fleets, for example heavy but slow lowrange fleets which act more like a homeland guard on sentry, heavy attack fleets with reasonable range + speed that are maxed in logistics to clear out enemy defenses and bases, standalone superfast interceptors that aim to attack non-military or nondefended craft.

when an AI gets bonuses that humans don't, it's not as hard to make it a challenging adversary. it can even be better than "perfect" if perfection is defined by how well you can do without any bonuses.

i personally wouldn't rate anything that doesn't play fair too highly, but it can still be fun to play against as long as the cheats/handicaps are designed appropriately.

i do think that the personality and character of opposing empires becomes more important the farther away a game gets from being fair/symmetrical because it's kind of insignificant to appreciate the difference between a 50% discount and a 60% discount

One of the things that Sorian AI did really well was provide a better AI at the cost of fewer CPU cycles which eliminated slow down on CPUs of the era. More bang for the buck and more AI expansions without losing gamespeed due to simtime calcs. Anyway the Gal Civ Strategic AI maybe ponders a lot, but it has been said they don't build a planet properly right now which should be step 1 of sorting out the overall production and military fielding. You need to have the industrial base to create the military force. More than thinking this should be a simple series of rule, does a planet already have a farm? Yes, then don't build another one, unless size is planet type 20-25, etc.

The AI fleet planner should also be building to maintain a force based on a few different states too. Peace, pre-war, active war-winning war, active war-losing war, where each state indicates how aggressive the AI is pursuing production of the military. When you are winning a war you don't need to build more force as aggressively as when you are losing, when in a pre-war state it makes no sense to build as fast as possible and when you are at peace with no enemies it makes no sense to build aggressively either.

OK, feeback from latest game on challenging:

Yor dominated most of the game, they should have wiped me, but I climbed back and won.  I was Alterian.  They had 4x the power that anyone else had, but I was able to win. Here's why:

1) AI did not adapt to my ship designs, nor adapt at all.  It was in love with missile boats with no shields.  So I went all beams, and all point defense.  I didnt have to worry about more rounded defenses until the AI got carriers.  Yor finally started to have some beam ships, 100 turns later....

2) AI did not seem to use its tech wisely.  For most of the game, until I went unto hyper-drive research mode, Yor and I had the same number of techs, but it seemed I had the tech edge everywhere.

3) Even DURING WAR, the AI at times would send in a lone ship very deep into my territory, even though it was encountering heavy losses throughout the galaxy by me.  Sending in probes might be ok to test your enemy, but it needs to have MEMORY, to know its not a good idea next time. I was NOT AT ALL sneaking behind enemy lines, picking off ships headed to "safe" rally points.  haha, on a few occassions, the AI in fact caught a few of my lone ships It almost like there was an AI Rally Point in the area, because these lanes were longstanding battle grounds for MANY turns.  Its like the AI didnt realize that I secured dominance, and all of the other friendly fleets in the area were wiped clean.  Maybe these were probed to look for my rally points....but I killed 10:1 in this scenario.  Yes, the AI caught me with my pants down a few times, but not worth all of those lone ships.  Safety in numbers when in enemy ZOI, and especially when in enemy ZOC.

4) When I finally started to push, I found several stacks of ships, NOT joined in fleets.  A few, but many single ships which were even more easy to destroy. These were sitting right next to planets.  I found this on a few different planets.

5) AI needs to learn that miniaturization and fleet logistics is OP.  Players do it, so should the AI.

6) AI needs to learn when to take a defensive posture, rather than kill kill kill. This should be based on performance over the last X turns.

7) When the AI was really pushing hard, I turtled pretty fast.  They were unable to invade the 2nd planet, as I rushed planetary defenses, military academy, buffing with adjaceny bonuses, rushing approval buildings, and trading for tech to defeat their WMD choice (BIO, it was VERY effective on the first planet they invaded).  The did continue to push past my planet which they could not crack, however atleast 10% of their TOTAL EMPIRE WIDE SHIPS just stayed at that planet for ALL of their offensive, and only moved once I started to push into their territory elsewhere.  

8) Know when to stop attacking.  Again, based on losses, AI should realize their offensive isnt going well.  I killed 1k ships, and lost 100, but that would have been much less if I wasnt lazy when I knew it was EVENTUALLY going to be a ROFLSTOMP.

9) Know when your Military Stength is useless, due to technology gaps.  Quite honestly, if the AI threw EVERYTHING it had at me simultaneously, I think I would have still prevailed, but BARELY.  Again, the AI felt it was "strong" and I was "weak" due to the number of ships.  But when I have better tech, huge logistics, and miniaturization, an incredible ship design mismatch, it just doesnt matter.  AI should have put more focus on tech, and less on ship building.  My god, at one point I was outnumbered ship wise like 20:1.  But they were shit ships, and shit fleets, compared to mine.  

10) And honestly, I think I would have lost if this hadnt happened.  I was princess lea, crying out for ObiWan.  I had to bribe the other races to declare war on Yor, in my darkest hour.  I needed the heat taken off of me.  Now Yor had 3 enemies.  But 2 enemies were a cakewalk, but a 3rd enemy kept on killing all of his ships, rarely killing any, but damaging them.  Hmmmmm....AI should have realized to go after the TRUE weaker opponent at that time (the other 2 AIs) and hold ground, regroup, get better tech ships, ones to counter my ships, etc, and NOT attack me.  Suspend the offensive as it were.  Well, it didnt.  It split its forces and attacked all 3 of us.  The other 2 AIs surrendered to me.  Yes, their planets helped, but Yor should have easily conquered/invaded most of those planets, I was too far away to try to help.  But it didnt, and I reaped the benefits which doubled my own power.

Lots of improvement can be done in the AI, but I think if logic was added to address most of the scenarios mentioned above, it would be MUCH harder than it is now on any given difficulty level.

L.     did you miss the first line ? 

Dansiegle,

loL  What a hoot,   Your write up sounds like you were looking over my shoulder taking notes while I played.