Announcement

**The diplomat** · February 4, 2000, 16:09

I think that civ3 needs to dump the intuitive AI (ie the AI type where the computer is given basic rules of what to do and not to do) and adopt an analytic AI, the kind that solves problems. For example, there is a game in development called DEFCON that will use an AI called Evolutive Human Emulator (EHE), which uses neural net theory.

From the DEFCON website:

"Defcon's artificial intelligence core, the Evolutive Human Emulator (EHE), is a revolutionnary use of the neural nets theory and the expert systems theory. It was constructed to behave as realistically as possible, to be less predictable and to adjust to its environment and from its experience.
It accomplishes that through thousands of simulations each turn, all selected according to past memory and through data analysis. Instead of using a standard «if... then...» chain, we approached Defcon's EHE through a problem-solving chain. The difference with the EHE is that there are no solutions given ahead in the code. The EHE always figures out by itself the best solution for a given problem.

We won't explain here the theories behind all this, because others did it better than we could. You can visit these pages concerning neural theory and expert systems.

However, we would like to point out that we adjusted these models to build the EHE, and that we added a couple of human behaviors:

A country can become frustrated by a problem, and start behaving more and more aggressively.
It can think ahead, and plan the response.
A country has favorites, that is other countries it will call first
A more developped and democratic country will evaluate the possible effects of its actions on many other problems, where a dictator might not care as much and go to extremes.
Not only does a country react to a situation, but it also tries to prevent them by dealing with possible problems."

Permission is granted to quote, copy, or otherwise reproduce the materials on Anonymes Informatique's website, provided that the material is not modified and that the following copyright notice is retained on each and every copy :

Copyright © 1997-99 Anonymes Informatique - All rights reserved.

I think that this is the kind of AI that civ3 needs. Not just to reduce micromanagment but to improve the AI's behavior in diplomacy for example.

------------------
No permanent enemies, no permanent friends.

**The diplomat** · February 4, 2000, 16:10

I forgot to mention: the DEFCON website is www.anonymes.com/defcon.html
[This message has been edited by The diplomat (edited February 04, 2000).]

**Theben** · February 5, 2000, 10:50

I think anything that reduces micromanagement, especially in the late game, can only be a boost to playability. I also think that Firaxis could allow the player to alter the game w/in the game files as to affect the AI- it's highly likely that some player out there will eventually come up with the right amount of tinkering to make a truly challenging AI. FE, be able to change:

-Starting locations of AI (in a better place for expansion)
-Wonder effects
-Expansionism, depending on it's usefulness in civ3 (you can "kind of" do this in civ2)
-Player can "tell" AI the best units to create, how many it should stockpile before sending into combat, etc.
-Player can "tell" AI how to set up cities for max production (in both game files & as an ally in the game)
-Never let the AI trade away a city with a wonder (as SMAC *sigh*)

As to micromanagement:
-Player queues in the files for terraforming priorities, which can be changed in the game
-Also have queues in files for the AI which the player can improve as needed
-Allow cities in late game to appear with some city improvements depending on tech
-Plus I like my idea of allowing for independent creation of farms, basic roads by your population

Good luck with your game btw. Your not trolling for ideas here again, are you?

**NoviceCEO** · February 5, 2000, 17:01

I'll support you all on this. I don't have as much experience on Civ-style games as you do, but I have the excessive micromanagement and the dumb AI. I played only one Transcendense level in SMAC to the end. The reason is, there is a point in the game that you have dozen of terraformers and you have to give them orders, which in all is a smaller problem. There's nothing as annoying as a reading reports from over 80 cities and selecting productions in a city that will take 50 turns to complete anything.

But the worst situation is when switching government types when that will give you drone riots in many cities (even if it was just close to be ok, you only needed to switch a Librarian to a Doctor). First, you receive 80 messages telling you that you have riots, when all you really wanted to know is about your ten best cities. Then, after you correct them, in the next turn the game will tell you that all those cities are OK, and you have to check them all. If you don't, you may miss something really important in the report screen.

One thing that could be helpful is simply getting a better Report screen and it's not that hard, it's quite simple. You would have a Report Screen that opens and take more space in the screen. A bigger screen would be easier to read, than that tiny box in SMAC. That screen should also have different colors for some words, to give emphasis on the cities that are more important from you. The entries should also be easy and fast to delete, using a one-click system (those pop-up menus really drove me crazy in smac)

In SMAC, I usually spent half or more of the total time simply reading the Reports Screen. Making it better would simplify things and let the game more enjoyable. I think we may not need revolutionary concenpts in the game (even though it might be necessarry in some areas), but some nice tweaks that would make the game better.

I think that smac was a title that was released without being quite finished. Maybe because Firaxis just rushed to release it or simply faced problems that it couldn't fix. I wish that doesn't help to Civ3, but anyone else has some more simple tweaks?

novice

**Adm.Naismith** · February 5, 2000, 19:47

I really don't feel the need for a user programmable AI, because I'm afraid it's really a complex task.

Because I really support your idea of some improvement into AI, with some autolearning model, may I suggest Firaxis build the AI so that we can export the knowledge database and
cross them with other players one?

I mean, if some player are very good to challenge the AI, it can learn more triks than playing against me. So, if I want a better challenger, I would like to share the AI knowledge of (by example) famous player as Zso Zso, Velociryx and so on. They will probaly proud to share by FTP site (Firaxis' owned or fans' owned) an export of their pet.
If I can download it and cross it with my AI (e.g. with few session of blind autoplay between the two models, if simple crossing isn't viable) I can end with some strong enemy and, more important, with a personalized enemy, a feature no actual game can offer.
Of course it's not an easy task to program, but can really become a "killer model" into the game market!

Of course any polishing to the game interface can help to make micromgmt more easy to live with, still I think than more effort into macromanagement area and AI able to learn will pay a better return of investment with CIV fans.

------------------
Adm.Naismith AKA mcostant

**The Joker** · February 6, 2000, 08:44

I REALLY like the idea of having an AI that learns. It would be really cool. If Firaxis could make Adm.Naismith's idea possible the game would truly be great.

**Mark_Everson** · February 6, 2000, 10:18

Hi everybody:

Thanks for all the responses. I've been off playing Multiplayer civ for a change! Let's see...

The diplomat:

The AI you cited looks interesting. A solid "expert system" approach would help out civ a lot. I'm a little skeptical about coupling neural nets with expert systems in this context. I am very familiar with neural nets and their advantages and limitations from work. The problem with a neural net for something like this is that neural nets require "training" on a very large data set to produce decent results. The data set not only generally needs to be fairly large, but it also needs to effectively "span" all the situations that it might have to respond to. For game as wide-open as a civ-like game, with endless combinations of strategies possible, it seems very difficult to me for them to actually get the appropriate training set. You can't use just machine-generated strategies, because you might end up with an AI that had a glass jaw with respect to several types of human-generated strategies that the AI didn't come up with. I'm not saying it can't work, I'm just very skeptical.

Theben:

I'm Always trolling for ideas Any time you'd like to give us some ideas, check out the Clash web page to see where current discussion is, and let us have 'em.

Your list on AI requisites is pretty good, but it really needs some decent "infrastructure" to back it up. By infrastructure I mean things like true understanding of the map, and positional realities. (If you want to see my opinion on a relatively easy way to do this, check out the mapAI link off the Clash web page) The adjustments you talk about will go a certain way toward improving things, but they can only go so far without solid bedrock to stand on. My contention here, is that the AI cannot make the correct decisions unless it knows the operational realities with respect to each potential foe or ally. But letting someone tweak the AI would, as you say, produce large benefits in and of itself. I think the main problem with the AI is that the designers really have no idea how to play their own game well!

I'm trying to thing with the people creating their own farms etc. in Clash, I think it will turn out great!

NoviceCEO:

Strongly agree with your gripes... The problem is, like I said above, I think the designers really have no idea how to play their own game. I think that in playtesting they probably only used 10 cities or so, because that is sufficient to beat the Awful AI. Then come in players with higher standards who want to push the pedal to the floor and see how fast they can advance, and the micromanagement becomes intolerable. I can't say too much about SMAC, because I played the demo once and was Shocked by how little they had improved things over civ. So I never got it.

Adm. Naismith:

There is a game that takes the "copy expert players" approach for the AI. It's called Galactic civilizations... it was an OS/2 game way back when, and I've never played it. But the reviews on the AI were pretty good. There is a new version coming out for Windows that will hopefully be out within the year. If you check out the Alternative Civs section here at Apolyton, you can find a link to it. I think that AI approach will only get you so far, but "so far" is a lot further than what they've been able to manage in civ!

Unfortunately, I don't think they're going to listen to us diehard fans. The clamor for good AI, which would IMO Revolutionize the game, seems to be lost in the din of requests for other features, which are to my mind just rearranging the deck chairs on the Titanic .

**tobyr** · February 6, 2000, 12:10

My two cents:

(1) Neural net logic would be nice, enough to let the AI learn not to repeat its mistakes. One of the most boring aspects of civ2 is watching the AI do something that does not work, over and over again.

(2)Expand the events macro language A LOT to let people tinker with "gifting" AI tribes to improve their actions. Add a lot of hooks to the software where events can occur, such as "population increase to N", etc., etc., etc.

(3) Try to offer a scripting language or library API to let people rewrite or expand the AI logic. In the long run, hundreds of players can do better than Firaxis, or any company trying to make a profit, to improve the AI.

(4) DON'T make civ3 much more complex than civ2, because the AI gets MUCH harder to improve as the game gets much more complex.

(5) Build a PALM PILOT VII interface so that people can play MP from their hand-helds.

- toby

------------------
toby robison
criticalpaths@mindspring.com

**The diplomat** · February 6, 2000, 15:08

Mark: I'd like to respond to a valid point you made where you said, and I quote:
" The problem with a neural net for something like this is that neural nets require "training" on a very large data set to produce decent results. The data set not only generally needs to be fairly large, but it also needs to effectively "span" all the situations that it might have to respond to. For game as wide-open as a civ-like game, with endless combinations of strategies possible, it seems very difficult to me for them to actually get the appropriate training set. You can't use just machine-generated strategies, because you might end up with an AI that had a glass jaw with respect to several types of human-generated strategies that the AI didn't come up with. I'm not saying it can't work, I'm just very skeptical."

DFiraxis said that to check if the new factions in SMACX were balanced, they ran Alien Crossfire with all factions computer controlled. They just let the game run by itslef, non-stop.
Firaxis could set up a game with different basic situation (strong civ, weak civ, sea civ, landlocked civ etc...)and just let it run for a weak to build up a database. With a million games under its belt, the AI should have a decent database to work with.
Furthermore, the advantage of this kind of AI is that it continually learns. When the human player, plays a game, the AI could keep adding to its memory, and learn better strategies, not just computer simulated strategies. They could probably have the Ai learn from multiplayer games as well, adding human-human strategies.

I'm like everybody else: I want a good AI for civ3. This type of AI intrigues me because it can learn and might be the closest thing to an AI with human-like strategies.

------------------
No permanent enemies, no permanent friends.

**Sikander** · July 1, 2000, 19:31

Diplomat,

Unfortunately, the letting the AI play itself method falls far short of what is necessary. Sure you get a large data set, but that data is crap. Compared to the ability of even one human (not to mentioned numerous intelligent humans linked together on Apolyton) it is insignificant. What is necessary is a data set of many human games, and a very powerful computer or set of computers to do the analysis. Perhaps if the game kept a record of every decision, and then uploaded the data to the host a proper human influenced data set could be created.

With a game as complex as Civ or Smac, and such a huge data set, Fireaxis (or whoever) would probably need to use distributed processing to chew through the data. Perhaps volunteers could leave their computers hooked up to the internet while they are away from them, and the host could send bite sized chunks of analysis to be processed.

Consider that the simple game of chess (which is many orders of magnitude less complex than CIV or SMAC) took years of effort for the AI to master. This is an industrial strength problem, and it calls for gross simplification, or an industrial strength solution.

**The diplomat** · July 1, 2000, 21:15

I suggested a learning AI because it is so intriguing. Whether it works or not or would be the right kind of AI for civ3 is another question. Your comments are well taken.

For civ3, an open AI, where each player could improve the AI through a .txt file is probably the best practical solution.

I definitely think that the AI needs to be fixed so that it does not do stupid things like moving settlers in circles, or attaking a city over and over again with just one unit.
However, let's not forget that a trully unbeatable AI is IMPOSSIBLE?

------------------
No permanent enemies, no permanent friends.

**UltraSonix** · July 3, 2000, 04:24

I agree that an open source AI is probably the way to go, but the user-alterable bits of the AI would not be the actual engine, but rather the scripts. The scripts would say things like:

science rate:50
tax rate:50
luxuries rate:00

If land is large:
build 5 city, defend with 2 militia each.
research goal monarchy.
switch gov to monarchy.
build 10 city, defend with 2 phalanx each.
...
Stop when reach gunpowder

If land is small:
build 3 city, defend with 1 militia each
research goal navigation
build 2 trieme
research goal republic
build 3 city, defend with 1 militia each
...
Stop when 15 cities

This is obviously only an example - I'm not an AI programmer, so I don't know what goes into doing something like this. But I feel that a game like civ3 would not be able to have a good enough AI that would learn. So having a scripted opening seems the next best option.

------------------
No, in Australia we don't live with kangaroos and koalas in our backyards...

**The diplomat** · July 3, 2000, 08:54

I have come to the conclusion that micromanagement is the problem. It is the reason why the AI has problems. Therefore, to have a better AI, we need to reduce micronanagment as much as possible. Templates are not enough. I am not talking about governors, but about changing areas of the game to make them more AI friendly.
Here is an example of what I am talking about:
Some time ago, an idea was offered to replace spy units altogether with a spy screen ala botf. I believe that a spy screen is fundamentally more AI friendly than spy units. Here is why. It is very difficult to teach the AI how to be smart with units. (the vast majority of the complaints in Chris Pine's AI thread are about how the AI handles units)
Once the AI has determined that it needs spies, it needs to:
-determine which city should build the spy unit.
-find the best target and mission.
-find the best path to that city.
-deal with enemy unit encountered on the way towards the target.
That is four problems that the AI needs to solve, in order to excel at spying.

The spy screen model requires the AI to:
-determine the need for espionnage.
-determine the right target/mission.
This model only has two or maybe three (if I missed one) problems the AI needs to excel at, less than the units model.
The model is more AI friendly because it has less variables for the AI to handle and it eliminates the spy unit which is such a stumbling block for the AI. Seeing how difficult it is for the AI to intelligently move units around, just replacing the spy unit would help a lot.

Now, I am not suggesting that civ3 should eliminate all units. Of course not!
I am suggesting that if we really want a superior AI, we should look at areas of the game that are too difficult for the AI to master, and change them to reduce micromanagement and MAKE LIFE EASIER for the AI.

Some of you might say: "Are you suggesting The diplomat, that civ3 should be completely changed? That would make it unciv like!"

Would Public Works, instead of a settler terraforming, make the game less civ?
Would a spy screen instead of spy units make the game less civ?
Would a trade screen instead of a trade unit make the game less civ?

I am not suggesting the game should be completely different. But I think certain areas could be changed to make life easier for the AI.

Why make life unnecessarily harder for the AI?

------------------
No permanent enemies, no permanent friends.

**Ralf** · July 3, 2000, 09:40

In response to above post, by the diplomat:

Your reductional approach to what the AI should mess around with is really bull’s eye. I agree totally.
Lets just hope that Cris Pine read this, and take it to he’s heart.

I am usually against cross-posting, but you should really post this in the AI-questions thread as well .

[This message has been edited by Ralf (edited July 03, 2000).]

Announcement

Civ 3 Needs Much Better AI - would help with Micromanagement too...

Civ 3 Needs Much Better AI - would help with Micromanagement too...

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment