Hmm, dug up this nugget from the old threads. From Frogboy:

"Some of you guys are losing track of the other point of the metaverse -- it's supposed to be a vaguely accurate way of determining skill."

It's really vague now though.

Hey, does anyone remember the thread whereby Brad mentioned he did work on Warcraft or something? I seem to recall him saying that he was the 3rd Human player on the WC ladder or something like that.

If the Metaverse is supposed to be a vague way to determine skill, then the scoring differences between difficulty levels should be quite extreme. Right now, as has always has been, simple games on larger maps outdo difficult games on smaller maps a lot of the time. The obvious conclusion from observation is that the Metaverse rewards commitment above all else. Just turn in games, the more the merrier.

On the good side, I am glad to see that scoring has finally been levelled out and standardized. The continual score inflation was real irritating.

Ray--

Thanks for the insightful discussion, definately some things to think about.

1) I am also suprised you have played a single title for so long. the metaverse and ranking is almost certainly what has given you the incentive-- good design on Stardock's part there. I tend to tire of games after about 3 weeks as they all seem the same to me these days. This of course has something to do with the fact that the average cost to produce a AAA title is $2 million USD. Its so pricey that publishers will not underwrite original ideas; they want a product with a proven revenue model.

2) Today's hand held chess machines are so powerful that they will even beat the masters. I know two Chess masters personally, and they consitantly lose to the highest skill level AI on hand held chess machines. YES THEY ARE THAT GOOD. No supercomputer needed. You have correctly pointed out that there are fewer variables in chess to track, but I think the real reason that chess AI is so good is not that it is not a complex game, but rather we have had hundreds of years of playing data on this game to perfect strategies etc. I have a feeling if we played galciv for 200 years slowly tweaking the AI it would be formidable indeed!

3) You have asked for Examples of Superior AI in computer games. Again I return to chess (Sargon series for instance) and Warcraft III. If you have played Warcraft III in the campaign mode, you have not played against the expansion AI. The campaign mode for warcraft is designed to teach you the interface, the units, and the basics of the game to make you ready for skirmish mode and multi-player. It uses an attrition AI combined with spawning units in campaign mode. I can see how you could confuse this however, as GalCiv HAS NO campaign mode. It drops you immediately into skirmish mode, no campaigns or tutorials. Now do this for me: Load up warcraft III and run a skirmish game for yourself versus the hardest AI. Just 1 X 1 for now. Your gonna lose. After you do, WarCraft III Has a totally Awesome feature where you can watch a playback of THE ENTIRE GAME. Watch how the AI beat you, it did not cheat, you should learn something from this.

4) Based on how fast the metaverse submissions take place it cannot be uploading enough data to really clone a particular player. I agree with your more realistic vision of what is ahead. I do not think that replay value is only added by making a game "tougher". Tougher assumes that you gain principle satisfaction from a game for score, or the win, not the actual game play experience. I think replay value cold better be added to galciv by enhancing that gameplay expereince. One of the best ways would to be to make the non-military victory paths more viable to encourage a more flexible and less focused approach to vitory in any given game, based on how the individual game unfolds. Just out of curiosity which computer strategy game, in your opinion, has the greatest replay value of all time?

5) Now the fact that you can win 90% of your crippling/ Maso games with cultural victory I find VERY INTERESTING. That also gives me a good deal of hope. At some point when you have free time do you think you could replay my WOTR challenge and achieve cultural victory in roughly the same timeframe and repost your AAR?

6) Achieving a balance where each victory path has approximately the same chances is not a dream state. It can be achieved. It is however an iterative process which must be reached through trial and error. If the consensus is that military victories are too easy at present, then you tone them down a little and burn more games to see if the distribution comes out more equal.

7) Most strategy games with scoring use time as a detractor, aka you get more score the faster you finish, but this is balanced with all the other score values. Surpised to hear you say glaciv does not take this into account.

8) The Galciv AI is not different than the MOO2 AI. The MOO2 AI does not cheat until you go above the normative skill level (normal) to hard. Galactic Civ does not cheat until you go over the normative skill level (intelligent) to bright. Personally with my limited play experience I have found the sub-bright AI to be plenty of challenge so far. I have a feeling my interest in the game will top out at the Intelligent level of AI.

9) If Brad meant MASO to be unwinnable then I think he wasted a great deal of time that could have been spent on other development issues. Even a novice programmer can create an Akobi-Ashi-Maru scenario that will beat the best player in the world every time. If you are going to make an AI it should be beatable, even if it is VERY VERY tough. It should however have a good chance of beating the human player at least part of the time.

10) I agree with you that I need more hours under my belt before I jump to any conclusions. That is why I labeled this thread "concerns" not "problems". It sounds like the end cycle is not as gloomy as I was worrying so I will spend some more of my precious time on this title without guilt.

Pontius--

I agree that chess is a more simplistic environment than GalCiv. Another big difference you did not mention of course was the random factor in GlaCiv, none of which exists in Chess. Yet chess is the basis for all AI.

I am well aware of what an AI is, but I am also well aware that the goal in writing a good AI is to be as human as possible. I think your programming skills may be a little dated if you think it can never exceed the sum of its conditionals. Years ago that was true, but everything changed with the advent of neural net and fuzzy logic. These platforms actually allow a computer to LEARN independent of any hard wired conditionals. Thanks again for posting.

Wow, the forum is bustling with activity.

ThelebK, your strongest point is in the Warcraft 3 portion. The AI is indead very good at playing Warcraft 3 and the toughest AI can defeat the vast majority of the players out there. The difference here is of course the difference between RTS and TBS.

In Warcraft 3, the AI faces a small tech tree, small number of stategic points, and one sole source of income. It also has the advantage of being able to handle everything at once.

It memorizes a certain number of patterns and runs the entire game with that. Not only is it fast, but also better than most humans can pull off.

But you'll notice that aside from the build and expansion the AI does not run any particular part of its game particularly well. Its assault against creeps is relentless, but loses quite a few men in the process. Its micromanagement is mediocre at best, but Warcraft 3 is a game where having any micromanagement is relatively good. Its attack is disorganized and tactically poor. Most of the time, the attack is a long string of attackers, not walking in formation, and takes damage needlessly. And their choice of attack is downright stupid.

The problem in TBS is that the AI doesn't even have the advantage in build up and expansion, and the millions of permutations of star system locations make it impossible to identify the sectors that have true strategic value. And like in Warcraft 3, the AI doesn't send in a coordinated attack, but rather a string of ships at the target. But since battles are swift and death dealt in one single turn, that string of ships is worthless. The larger tech tree requires significant long term planning, but emergencies and sudden changes requires short term management.

But I'll have to agree, an AI with a good command of tactics and strategy would make for an excellent game. An AI that would know how to counter a rush, destroy an alpha strike, and run you into the ground economically and culturally, that would be awesome.
[Message Edited]

Oh, yeah. Right now the easiest path to victory is cultural victory. Domination of 50% of the star sytems on a map and a few influence resources can get you there.

Since military victory scores significantly higher, most players turn off cultural victory so they can obtain the military victory. I've seen about 50% higher scores for military victories.

Hi White Council folks,

I thought I'd peek in on your thread and I saw your discussion on AI quality. Warcraft III is repeatedly cited as an example of "good AI."

It seems that none of you are aware that the toughest level AI in Warcraft III is not tough because it's so smart. It's tough because it cheats--it gets double the resources. So a peon for the human player brings back 10 gold per trip to the mine, whereas a peon for the AI gets 20 gold per trip. And likewise for lumber, I think. This is obviously a huge, huge advantage.

I think you probably could make a Warcraft or Starcraft AI that could regularly beat a human, but that would mainly be due to the real-time elements--a computer can micromanage certain things much better than a human because it thinks thousands of times faster. For example, imagine playing Starcraft against a computer that treats every SCV as if it were a medic for its machines, so you'd have dozens of repairing units all working together to support a tank push. (I hope at least a couple people understand my Starcraftese.) There's just no way a human could click fast enough to imitate such a strategy.

[Message Edited]

Weekends are usually quiet, Maike but it seems especially so on this long holiday weekend for many Galcivers.

Welcome to the White Tower, vincible. Our doors are always open to guests such as yourself.

The real time element is the main reason why I prefer TBS to RTS. I'm not a fast-clicker and some part of me resents the fact that there is idle capacity lying around or soldiers idling because I'm not a fast enough clicker.

Imagine if Chess was realtime instead of turnbased

There is logically no harm in creating a level that
cheats so long as the non cheating levels are as robust
and challanging as possible. Just dont play the
cheating levels. I have never seen an Ai in a tbs 4x
game that was remotely challanging without cheating.

I dont even think I have ever seen an Ai that could
come close to beating me in a tbs 4x game without
cheating the very first time I played the game.

I know its possible with the technology we currently
possess to create such an Ai with very advanced
learning machines, I have seen some awesome potential
in what is coming out in the Ai theory department, but
the development of that Ai is way beyond the means
of any gaming company for several years yet and more
importantly, the running of that Ai is ridiciously
beyond the means of the computers that people are
playing such games on.

It should be noted that the way such AI's will be
developed is by a method much like the metaverse pretends
to be. By playing a large pool of players and learning
patterns through massive trial and error.

But as Thelebk noted there is no way in hell the
current metaverse is actually doing that.
There are 2 things that are obviously missing from a
user point of view.

1. A massive learning pattern through trial and error
requires rather quick feedback, where the AI is actually
updated regularly through continious generations of
advancement. This would need to be automatic and
fairly quick, as in updated everytime you start a game.

2. Currently the AI is updating the tech tree path
through manual adjusments by the dev once in a great while.
This is very flimsy and doesnt even touch upon the AI's
actual weakeness which is tactics and diplomacy.

I have seen alograthims for learning logic and I have
to admit the potential is actually awesome, this isnt
to be confused with awareness, which is an entirely
different subject, but given the right conditions
and the right learning alograthim, a computer
can learn most anything.

Computers learn by having an encoded logic, for
comparision, call this the DNA sequence. Then several
versions of the AI try their best to acomplish some goal
In this case, beat a human player. Those that
did better are then mated in a semi random way with
other succesful AI's. A new generation of AI's are
then created through random combinations of the previous
succesful AI's and the process starts over again.
through hundreds, thousands even 10's of thousands
of itterations, some very surprising things can occur.

They have used this logic to create electronic circuits
that are better then any huamn designed circuit. In
fact AI designed circuits have been patented and put
in production.

Some of the designs have even baffeled human experts on
how they even work, but they clearly do.

There are actually several types of the same logical
development that have been used, including one
that simulates the evolution of walking in a simulated
world as well as that simulates a device that learns
to follow a light source in the real world. The neat
thing about this last one is that it actually developed
an unexpected split in the individual system logic
based on wether it was day or night in the lab the
machine was being run in.

I could find references to such things if you wanted,
but it is possible, but it isnt being done here but
it might be done in the future, though probably a
decade away yet before such logic could be applied
to something as complicated as a computer game, though
I think it will as it will be a natural extension
of the development of such AI(from the perspective
of those interested in AI development as a science, not
as a game)

Very interesting stuff, but galciv is not it, maybe
galciv III as I hope galciv II is multiplayer

Actually it's been quite noisy in this thread.

Not so elsewhere.

Thanks for your thoughts, but I am not sure I agree with you. By altering the economy you effectively are cheating at the three items that build off it

Yes but civ cheats on the economy, then cheats on the other things too.


Also you took my comment about Civ III out of context. No where did I state that CIV III has good AI. What I said was that the cultural model of CIV III had a much more prounounced game effect than the one in Galvic, and for that reason the cultural model, and only the cultural model, of Civ III was superior

In what way? Are you confusing culture in civ3 with nationality? Culture was fairly lame in civ3. You could see 1 or 2 cities flip on that but that's about it. You can flip the entire map here. it's an integral part of the game.



The invincible phalanx cheat (otherwise known as the phalanx that sinks the battleship) in Civ I does not work exactly as you think. When I talked to Sid Meyer about it at CGDC he explained that he used a sliding adjustment based on the power of the two nations involved in the conflict. Thus if you were winning the game and fighting a CIV that was doing poorly, you would have a big negative to your att/def #'s and the computer would have a big positive bonus. It had nothing to do with phalanx's; spearmen got the same bonus and could fight off tanks etc.

Sid explained it that way, but it did not in effect work that way. There were times when the defensive unit was made invinceable. And not just against 1 battleship. 15 in a row, 20 in a row, I even lined up 50 just to test. All 50 lost. As well as the 30 tanks and 10 artillary units. But if you'd reload and wait a turn, the invinceability was gone. Though the phalanx could still kill a battleship maybe even two, it couldn't come near to 50. So the invinceable phalanx cheat is seperate from the unit adjustment cheat he mentioned. The unit adjustment cheat is ubiquitous in all the civs save the two from activision (it has it's own cheats) Bottom line is it's still a cheat. A cheat that doesn't exist here.

Staffa: I would be quite interested in those references. I'm sort of into the subject.

http://www.abelard.org/turpap/turpap.htm
http://www.genetic-programming.com/
http://www.genetic-programming.org/

Genetic Programming: On the Programming of computers by means of natural selection
John R. Koza, MIT Press 1992

Genetic Programming: The Movie.

Genetic Programing: On the Programming of Computers by Means of natural Selection

Genetic Programming III: Darwinian Invention and Problem Solving

Genetic Programming III:Videotape: Human-Competitive Machine Intelligence

Genetic Pgramming IV: Rotuine Human-Competitive Mchine Intelligence

All by John R. Koza, Forrest H Bennit III, David Andre and Martin A. Keane

Ahhh MIT Press, I've seen some of the more advanced stuff they been doing with that. Nothing going with more advanced activities. Evolution takes a long time, and the process is very slow.

Genetic Computing on the other hand has some possibilities in getting an answer to the most controversial CS problem: The NP problem.

Guys--

Thanks everyone for posting to this thread I have found the constructive exchange of ideas very useful. Ok, will try and reply to all in one mega-mail.

Ray & Vincible-- yes one advantage an AI has in an RTS environment is that it can make thousands of minor tweaks all over the board instantaneously, and simultaneously, as it does not have to use an interface like a human does. This led to some pretty annoying things in Age of Empires as the computer could effictively "skirmish you" (aka atack with a long range weapon the run back when you followed, then shoot again when you stopped etc., all over the map at once. Used to drive me crazy heh. Vincible my play experience is a bit different with the high end Warcraft III AI. The last time I played versus two AI's it not only relentlessly got all the creeps on the map, it set its rally flags to its heroes, built multiple production facilities streaming units to its heroes, and then both AI's rushed me at once. (They also bolstered their forces by purchasing as many neutral units as they could at huts!) In its attack it walked right past my static defenses and went straight for my workers, killing them all, then my military unit buildings. I barely survived that onslaught. When I counter-attacked the nearer of the two AI's, his partner portal scrolled in and nuked me. Frankly I was impressed.

Staffa, yes I have heard about recombinant computer structures. It is yet another example of computing moving into an age where the sum of the whole is actually greater than its parts. This trend began with neural net and fuzzy logic. I do agree that the computing needs to run such an AI would be high end, but have we not already seen high end applications running by offloading a part of the processing required across all the CPU's connected to the system? Aka in this case the metaverse?

Pontius, I see what you mean about using culture for victory. But of course remember you do not have to flip cities to win by culture in CIV III you just have to get your nebulous culture score to a certain number, and no matter who owns what cities, or how many you have you instantly win the game. Now I am not proposing that for GalCiv, what I like about the Civ III culture model specifically is the way it counters the military rush which typifies most strategy games. By requiring high garrisons to maintain order it slows the advance of the military juggernaught. Newly conquered cities in CIV that are thinly garrisoned alomst instantly flip back to their original owner. In GalCiv it takes many, many, turns for this to happen rendering it a useless means of defense.
As to the invincible phalanx cheat what you are describing is clearly a bug, and not intentional on Sid Meyer's part. I am sure we can find bugs in GalCiv as well. If fact just last night I noticed that an AI player had sent a task force to take a planet of a mutal ennemy. I conquered the planet first and garrisoned it. When his task force arrived it did not acknowledge the change in ownship, suicided against my garrison, and went to war with me by accident. Interesting...

Different game mechanics. In civ you used military units to conquer and subjugate the population. In GalCiv you use your own population and kill enemy taxpayers. Doesn't make sense for your own pop to revolt against you.



That is a bug. Haven't seen it in a while, not since 1.05. Be sure to check your version and send it in.

Re RTS/WC3, I think its only fair in terms of competing with a multi-tasking comp that I can execute my commands based on my thoughts through a PC-gamer mind interface rather than my mouse-hand. But then it usually isn't so much about fairness isn't it, it is about fun and challenge.

Ahh, nothing like a scholastic discussion on the "intelligence" of electronic circuits here on a lazy weekend in the White Tower.

Looks like 1.09 has no effect on metaverse scores (EBZM)

And it has been a good discussion. Thanks ThelebK!

TelebK: what you cited as evidence of excellent Warcraft AI is nothing that the average human wouldn't do.

I asked this question on a Warcraft board. In addition to the resource cheat they suspected several other cheats. And in spite of this one person had managed to beat two allied comps. There may be more--it hasn't been that long since I posted the question. (I'd rather not link to it since it's a semiprivate board which gets a fair amount of spam already, but you can go to battle.net forums and ask if you like.)

that's some skill comparable to the best TBS players. Any cheese tactics? I don't think you could beat two AI players straight on.

Welcome to the Council, zyk0001.

Make yourself at home here at the White Tower. Feel free to pose questions or interact with our panel of residents and visiting professors.

Maike-- One of the tradeoff that is often made in computer game design is realism versus play balance. Since this is a game it is generally better to err on the side of making the game more balanced/fun than being hyper realistic. I agree that it is not realistic for your own taxpayers to revolt on a new world, but I am willing to suspend disbelief on this if it would add a new layer of complexity by allowing high cultutal scores to be a semi-effective deterant to the cheesy transport rush.

Vincible-- I agree that the Warcraft III AI in my recent game did not do anything a good human player wouldn't do. That's precisley why I praised the AI for I believe high end AI's SHOULD play like humans in order to challenge humans.

Maike again-- I cannot beat two warcraft III Ai's solo at the high end AI setting. Frankly I am glad I cannot as that would make the game too easy. As to which version of GalCiv I am running I will have to check versions numbers. I Installed the 1.05 bonus pack so I assumed that upgraded me in the process. Another bug I have encountered is the ability to build dreadnaughts without having the appropriate tech. With auto shp rebuild option on I conquer an enemy planet that was building one, and just let the que continue running and re-running...

More on Warcraft AIs:

Several people had beaten seven allied insane AIs using an exploit (the AIs won't expand if there's an invisible thing on the expansion). This exploit has since been patched.

Various people added that if you harass an insane comp early on then you can mess up the AI's highly optimized build order and after that the AI plays very weakly.

That's not due to influence. That is due to nationality and resentment at being conquered. Galciv doesn't have that concept.

Also, in civ3, when a city flips due to influence (a much slower process than galciv), such resentment along with several resisters that won't work untill pacified by the garrisonned units are not present.