AI longtime simulation feature

tesco samoa · #1 February 14th, 2002, 05:20 AM

I am not one to jump on a band wagon. But I know a good hoe down when I see one.

Fantastic idea.

# of turns would be good
Summary log would be nice as well. ( Ie the score page every 5 to 10 turns for all races)
It will be great for mod testing.

PsychoTechFreak · #2 February 14th, 2002, 10:11 AM

I thought about the stopping conditions Last night. Maybe this could be done by an extension of the victory condition setup, e.g. as a kind of victory/AI simulation stop condition. There are just a few disadvantages:

-the conditions cannot be changed during a game, means: you cannot simulate until one condition, change it, and go ahead to the next condition

-the conditions seem to be requested every 3 to 5 turns or so. I have tried to run a single human player (for weapons testing, ship setup and so forth), after the first few turns (it seems to be not always after the first turn), a victory window pops up plus a question about continuing the game

All right, I guess there are better solutions.

As for the additional stop/break simulation conditions, we have now:

1. War declaration
2. Hostile activity
3. Treaty breakage
4. Score point goal
5. Research area goal
6. First contact
7. Number of turns (of course)

Any additional ideas ?

I guess some of them are not possible or needed, but let's see, what else would be important for AI testing and for scenarios (dmm: thumbs up !)...

Shyrka · #3 February 14th, 2002, 10:52 AM

The idea is very good, not only for modders but also to MM. Why not "train" the AI to improve it just as neural nets, decision trees or other algorithms? (In these structures engineers run a situation and lets the system deal with it, so they learn from them errors. With time the systems take better and better decisions).

Q · #4 February 14th, 2002, 03:12 PM

Me too I would welcome such a AI longtime simulation possibility.

PsychoTechFreak · #5 February 15th, 2002, 09:33 PM

quote:
Originally posted by Shyrka:
The idea is very good, not only for modders but also to MM. Why not "train" the AI to improve it just as neural nets, decision trees or other algorithms? (In these structures engineers run a situation and lets the system deal with it, so they learn from them errors. With time the systems take better and better decisions).

Hey Shyrka, do you know "Fritz", my favorite chess program ? It has a self-optimizing engine which does not make the same error for a second time.

For SEIV this would be future music, but what a potential ! EDIT: "dreams of the future", of course. I am still thinking too much in german.

Regards from a dreaming PTF...

[ 15 February 2002: Message edited by: PsychoTechFreak ]

Shyrka · #6 February 16th, 2002, 01:50 PM

quote:
Originally posted by PsychoTechFreak:

Hey Shyrka, do you know "Fritz", my favorite chess program ? It has a self-optimizing engine which does not make the same error for a second time.[ 15 February 2002: Message edited by: PsychoTechFreak ]

Really? Then that's exactly what I was thinking. SE4 is so deep that for making the AI more effective some training algorithm is needed.

Any "Fritz" related web page?

PsychoTechFreak · #7 February 17th, 2002, 10:04 AM

quote:
Originally posted by Shyrka:

Any "Fritz" related web page?

Sure,

http://www.chessbase.com/catalog/product.asp?pid=85

I don't have the ver 7, this has been published just recently ,I have got ver 6, the older Versions are a lot cheaper.