Genetic AI Project

[The Great Apple]

Version 0.30 and generation 3 posted. See first post for details.

Note that there are now some datafiles with 50 AIs in in the third post. Use only if you're smashing through the 25 AI sets really fast.

 

[Gunner]

If anybody has anything about the AI that they find poor, and would like me to patch something in for I'm happy to accept suggestions. The simpler the better (like the promotions thing).

I don't know if this would qualify as simple, but to me one of the most glaring deficiencies of the regular AI is how it will always split its army up into individual or very small groups when its invading you. It effectively removes the AI's ability to utilize combined arms and allows the player to easily kill the small groups. I think one of the reasons why it does this so much is that the AI really puts pillaging as a very high priority, much higher than I ever do.

So to remedy this is there some way you could easily put the AI's pillaging priority down some? Even better if there was some way to tell it to keep its units stacked together.

Sorry if this isn't exactly the type of thing you wanted, but I figured it would be better to just throw it out there

 

[mike6426]

for number .3, the loading screen locks up at initializing, you can still hear the beginning track but you cannot see anything. This did not happen on .29 so what has changed that could make this happen.

 

[Craterus22]

The popup should come up at the start of every game (even after restarts) after the set is complete. Does it not?

True - but there is no indication of what iteration it is on while the game is running automated. You have to stop the game and re-start a new one to see if your pop up shows up (at least that is what is happening for me).

 

[The Great Apple]

for number .3, the loading screen locks up at initializing, you can still hear the beginning track but you cannot see anything. This did not happen on .29 so what has changed that could make this happen.

Anybody else had this problem?

I've changed a few things about how the AIs are loaded, and this is the sort of thing that that could cause, however I don't think this is the problem.

There was a problem with some of the earlier sets playing with large amounts of Civs - a few AIs would get trimmed and then there wouldn't be enough to load, and the program would get caught in a loop. I thought I had put something in place to solve this, but maybe it's not working properly.

is there some way you could easily put the AI's pillaging priority down some? Even better if there was some way to tell it to keep its units stacked together.

This should be encompassed by the genetics, but I can put in a few things that might encourage this behaviour to see if it is benificial.

 

[mike6426]

I did a test with four nations but it still stuck so it does not seem to be amount of AI, it rather seems to be an AI loading problem.

 

[klokwerk]

I have no problem with the 0.3. It works on my small pangea maps. I put aggressive AI on too, to see them fight more often.

Congratulations for the great work ! I wonder if it would be easy to make it work under warlords, because the Firaxis AI is much better in warlords and there are new concepts (vassal states, great generals) which are worth the change.

I wonder if we'll end up with a worthy opponent by randomizing all these parameters without altering more the basic algorithms. The second generations AI I already tested didn't seem to show there was a real 'dominant' AI. I mean most of them had games where they were highly ranked, and other games where they were wiped fast, so I don't know if we can get something interesting... I would like to see a true champion emerge in the next generations

How are produced the generations ?? How do you make them 'have a baby' ?

 

[Gunner]

.3 is working fine for me too.

 

[The Great Apple]

Congratulations for the great work ! I wonder if it would be easy to make it work under warlords, because the Firaxis AI is much better in warlords and there are new concepts (vassal states, great generals) which are worth the change.

It probalby wouldn't be too hard. A few hours work. I am, however, worried about patches. Not only would these change the SDK forcing me to do some merging, it might also significantly change aspects of the game (like the chopping nerf in Vanilla). I'm going to wait until at least the first patch, and may merge some of the findings from this into it - most of the AI is muchly the same, but they've upgraded certain areas.

I wonder if we'll end up with a worthy opponent by randomizing all these parameters without altering more the basic algorithms. The second generations AI I already tested didn't seem to show there was a real 'dominant' AI. I mean most of them had games where they were highly ranked, and other games where they were wiped fast, so I don't know if we can get something interesting... I would like to see a true champion emerge in the next generations

It takes time. Each generation pushes the AI forward a bit, and while there is a chance that a super-AI might occur in an early generation, the chance is very low.

How are produced the generations ?? How do you make them 'have a baby' ?

The AI changes from each of the datasets can be assigned a numerical value. Baby making occurs by both mutating (randomly changing some of the values up or down), or crossovers (mixing together values of 2 AIs). There is a filter so that the AIs who score highest are much more likely to mutate into new AIs, and be chosen for crossovers. The very best AIs (elites) are kept unchanged for the next generation, while still creating new AIs based on themselves.

EDIT: Post #3000!

 

[klokwerk]

Thanks for the answers and once again, congratulations !

Let's hope we see a champion AI some day !

 

[Surtur]

ver 0.29 AIData17 finished. Documents attached.

Noticed that Firaxis AI already had 8 games when I started. Is this intended?

 

[Arlborn]

ver 0.29 AIData17 finished. Documents attached.

Noticed that Firaxis AI already had 8 games when I started. Is this intended?

I had the same..This supposed to be tihs way?

 

[The Great Apple]

No. It was a glitch I didn't notice until too late. The thrid generation no longer has this glitch. Don't worry about it.

 

[se7en]

Is there a AI file somewhere with the top 8 in it to play against as a human? I don't ever quit anyway, I just minimize.

 

[Craterus22]

Is there a AI file somewhere with the top 8 in it to play against as a human? I don't ever quit anyway, I just minimize.

If we get a vote I would vote for a top 18 file every couple of generations (I usually play quick games with 18 civs on small maps)

Have some real life issues coming up so I will have to slow down my output, but here is a list of all the aidata sets I have sent in to date: 24, 39, 29, 27, 26, 19, 6, 8, 3 (twice?), and two (one each?)from genai version 25 and 26. Not too sure about how many I sent in for the early ones, but I have sent in at least 9 so far.

(had my pc working most of Labor day weekend - added in my laptop later on (the laptop performs too slow to play on but works fine for automated) - man if I was still in college I would have an entire computer lab running these things overnight)

I will be sending in ai56 on evening of Sept 6th...


ps - for the people stating that the Firaxian Ai is winning - what settings are you using? Do you mean winning each battle or highest fitness score? I am using mostly small maps (sometimes Standard, sometimes Tiny) and I don't think the Firaxian Ai has had the highest fitness score more than once (if at all), but then again I haven't been looking closely at that. I will in the future though...

 

[Gunner]

Here's one of the large sets from .30. Unfortunately Team Firaxis had the highest score

http://forums.civfanatics.com/uploads/50429/Gunner_s_AIData_30.zip

 

[Arlborn]

Here's one of the large sets from .30. Unfortunately Team Firaxis had the highest score

http://forums.civfanatics.com/uploads/50429/Gunner_s_AIData_30.zip

I saw a lot of ppl with Team firaxis haveing a high score, in my test that AI had a good score but not 1 of the best..But anyway what would it mean to the project exactely GreatApple? That the AI made by Firaxis still(at least for now) can beat the AI that you are makeing or you are as well changing the Team Firaxis AI?

 

[se7en]

I saw a lot of ppl with Team firaxis haveing a high score, in my test that AI had a good score but not 1 of the best..But anyway what would it mean to the project exactely GreatApple? That the AI made by Firaxis still(at least for now) can beat the AI that you are makeing or you are as well changing the Team Firaxis AI?

well, the firaxis ai was designed so its good to have it in there and it will probably always be a "decent" ai. the other ais are not designed yet (although i'm working on something to change that) so it would take some time to evolve a much better solution.

 

[Craterus22]

I saw a lot of ppl with Team firaxis haveing a high score, in my test that AI had a good score but not 1 of the best..But anyway what would it mean to the project exactely GreatApple? That the AI made by Firaxis still(at least for now) can beat the AI that you are makeing or you are as well changing the Team Firaxis AI?

I think he has stated that he is using Team Firaxis AI as a control AI - to the best of my knowledge that AI is not changing (except for maybe the mod changes that apply to all AI regarding the bug he fixed listed above on this thread on page 15)

well, the firaxis ai was designed so its good to have it in there and it will probably always be a "decent" ai. the other ais are not designed yet (although i'm working on something to change that) so it would take some time to evolve a much better solution.

Are you maybe working on a graphical interface to make it easier to modify the AI parameters in the AIdata files? If so, let me know how I can help, because that is something I was planning on working on. My hope was to create an editor to label all the 1200 values and let someone "build" an AI (either for their own fun - or for submission to The Great Apple's population for testing - perhaps with our own names attached?)

 

[se7en]

Are you maybe working on a graphical interface to make it easier to modify the AI parameters in the AIdata files? If so, let me know how I can help, because that is something I was planning on working on. My hope was to create an editor to label all the 1200 values and let someone "build" an AI (either for their own fun - or for submission to The Great Apple's population for testing - perhaps with our own names attached?)

Yup, it's coming along pretty well. I've got a working prototype. Hit me up on MSN: djprotocol7 at hotmail