Artifical Intelligence for World in Flames

A forum for the discussion of the World in Flames AI Opponent.

Moderator: Shannon V. OKeets

Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: abj9562
...
Still for location specific attack and defense values a possible solution might be to create a graduated extract of combat results tables of minimum defense and adding in additional units until max stack limits are reached. This could be extracted into a workable set of tables looking for optimum defense which theater commanders use to assist for specific location defenses.

However this would require the creation of a database containg all the combat/terrain variables and unit statistics iterrated through umpteen millions of combat results storing and averaging results resulting in the graduated tables. Quite a project in and of itself. So unless someone not under a legal hammer could take this on .... [;)]


As you point out, there are a lot of different factors involved in land combat in WIF. Here is a extract from a description of one of the standing orders:
================
The AI Assistant makes an educated guess (if fog of war is being used, a precise calculation can’t be done) of the following attack elements the enemy might bring to bear on the hex:
(1) basic land units,
(2) armor strength,
(3) invasion units,
(4) paradrops,
(5) HQ support,
(6) ground support (artillery and tactical bombers),
(7) defensive shore bombardment,
(8) surprise, and
(9) offensive chits.
=================

This doesn't take into consideration terrain or command conflict. In view of these, I think that rather than table lookups, procedures will be better suited to handle the diversity of elements.
Steve

Perfection is an elusive goal.
User avatar
Froonp
Posts: 7998
Joined: Tue Oct 21, 2003 8:23 pm
Location: Marseilles, France
Contact:

RE: Artifical Intelligence for World in Flames

Post by Froonp »

Unfortunately I can not help with your AI due to No Compete and NDA.
Hello,

I believe you've said either too much or too little.
Could you explain a little further ?
No compete with what ?
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Froonp
Unfortunately I can not help with your AI due to No Compete and NDA.
Hello,

I believe you've said either too much or too little.
Could you explain a little further ?
No compete with what ?

The words "No Compete" is short for a legal document that states the person who signs it will not compete with the other party. In this case, abj9562 is indicating that he has signed an agreement to not compete with another company's product - probably another game about WW II. The NDA is a non-disclosure agreement which is another legal document that requires the person who signs it to not reveal confidential information to people who are not part of the agreement. Taken together, what abj is saying is that he did similar work on another game and is legally prevented from working on the AI for MWIF.
Steve

Perfection is an elusive goal.
User avatar
Froonp
Posts: 7998
Joined: Tue Oct 21, 2003 8:23 pm
Location: Marseilles, France
Contact:

RE: Artifical Intelligence for World in Flames

Post by Froonp »

Well, in fact, I understood what he meant with the NDA, because I signed one before being on the CWiF alpha team, but I wondered what the other WWII wargame might be. Thanks you anyway Steve for the patient explanation [:)]
The curiosity of the playser I am is excited by the possibility that there might be another WWII wargame somewhere, that was it basically.
herwin
Posts: 6047
Joined: Thu May 27, 2004 9:20 pm
Location: Sunderland, UK
Contact:

RE: Artifical Intelligence for World in Flames

Post by herwin »

I'm currently working on an actor-critic AI for a model of an echolocating bat. The hard problem is that the state space explodes exponentially, so I'm investigating alternative representations--currently either sparse matrices or tree structures. I was not proposing that an A-C system be used for a single game, but rather it be trainable over a series of games at Matrix and the alpha/beta testers and then be frozen for release. It might be even enough to use it to develop an algorithm for player support to extrapolate from the current situation to the game end to see if resignation is in order. It needs to get the flavour of the Japanese Government estimating they had a 40% chance of holding America to a draw and thinking that was better than staying at peace.
Harry Erwin
"For a number to make sense in the game, someone has to calibrate it and program code. There are too many significant numbers that behave non-linearly to expect that. It's just a game. Enjoy it." herwin@btinternet.com
Sprocc
Posts: 3
Joined: Sun Oct 02, 2005 11:20 pm

RE: Artifical Intelligence for World in Flames

Post by Sprocc »

Hi Shannon

I gave up on WiF being finished a while ago - only became aware people were actually getting somewhere today.

I am very experienced in AI techniques. While, I'm waiting for my PhD final marks from an AI thesis, I also lecture KBS (amongst other things). My main areas of knowledge are KBS, ANNs and Reinforcement Learning. I have also been playing WiF since its early days (although not recently as my young kids tend to eat the units [:(] ).

I have had a quick look through this thread (so I am sure to have missed some suggestions) but thought I would offer some ideas. Also if you need any further assistance then please send me a priv msg.

Firstly, I am surprised you would attempt to develop an AI for WiF - It is the most complicated game I’ve seen with an absolutely massive state space. In my view the state space is your primary problem. To be successful you must reduce this through generalisation.

Certainly, the hierarchy of decision makers is the way to go. How these are implemented is the question.
I noticed one suggestion of using actor-critic (I’m assuming this was table based not ANN based as ANNs are too slow for this purpose). Even so A-C is still to slow, requiring many viewings. If using a table then the large state space becomes a problem. Also, unless you use something like CQL you are dealing with a single goal and getting enough training examples would be impossible for most uses. The main area I see for the application of A-C would only be in unit combat (possible including navel and air). Working out which would be the best amount/combination of units. Then you can apply a reward after each combat based n the outcome. This would be aided by the non deterministic nature of combat - where A-C tends to do better. Even given this - I feel a rule based system would be faster and easier to develop – and more importantly easier to directly modify later on.

My personal approach (and I know parts were discussed above) is a points based systems. However, each location with points would require more than one point rating. For instance it would need points for how important a target it is to take a point for how important to defend, etc and the points would need to change throughout the game. It would also be in a hierarchy corresponding to the decision makers. For instance, top level decision makers would have points according to larger based strategies such as countries/regions/convey lines/reducing enemy productivity etc. While the low level decision makers would have points attached to particular hexs etc (only in their particular sphere of activity - if moved to a different front they would get the points for the hexs in this new area). Each decision maker is then simply attempting to maximise points. Each decision maker would have its own set of rules to make a decision. In order to allow each agent to appear to do different things each time a small amount of noise applied to the point system is simplest and would achieve the effect of choosing different but good actions. The amount of noise can be based on the skills of the decision maker. For instance, an aggressive decision maker would tend to have positive noise applied to areas requiring a more aggressive behaviour.

Developing rules for behaviour of complex agents such as these would be incredibly time consuming and difficult. KBS in complex domains can be very difficult to develop with. They also can have difficulty generalising and if the engineer omits some information the agent can behave very oddly. To help in this regard I would suggest the use of a simple methodology called Multiple Classification Ripple-Down Rules (MCRDR) http://www.comp.utas.edu.au/users/bhkang/ .

The advantage of MCRDR is that it is developed through exceptions during maintenance. Therefore, it can be developed through actual game play. Testers play the game against each other they could be told what the AI would currently do - if the player does something different then the MCRDR systems asks why? The tester answers (by selecting differences in the current state space from what it had seen previously) are used to create a new rule (or sets of rules). Therefore, this method does not require people to just sit there trying to think up rules. It can be based on the way people actually play. If rules are based on general details such as points then the rules can be applied more generally – alleviating many of the problems from the larger state space. It tends to develop more robust systems and is well suited to this type of application. There has been some research into using MCRDR specifically for this purpose of game development.

My PhD was not directly in MCRDR but an add-on to MCRDR and could improve the KA process even further - However, at the moment my technique has not been released from an IP agreement so I can't tell you about it - if and when it is release I can detail it further.

As I have written this on the spur of the moment - I'm sure its wrong in places - its really just some thoughts.
If I knew what I was doing, it wouldn't be called research.
User avatar
Froonp
Posts: 7998
Joined: Tue Oct 21, 2003 8:23 pm
Location: Marseilles, France
Contact:

RE: Artifical Intelligence for World in Flames

Post by Froonp »

I don't understand all what Sprocc said [:D], but reading this, and also knowing that Steve has gathered around him a team of people like Sprocc to design AI, I fairly happy and quite confident that we are on the right way to have one day MWiF installed on our PCs. [&o]
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Froonp
I don't understand all what Sprocc said [:D], but reading this, and also knowing that Steve has gathered around him a team of people like Sprocc to design AI, I fairly happy and quite confident that we are on the right way to have one day MWiF installed on our PCs. [&o]


And I thought his post would be "intuitively obvious to the casual observer".

[I asked him to fill out a non-disclosure agreement with Matrix and join the group developing the AI.]
Steve

Perfection is an elusive goal.
User avatar
Greyshaft
Posts: 1979
Joined: Mon Oct 27, 2003 1:59 am
Location: Sydney, Australia

RE: Artifical Intelligence for World in Flames

Post by Greyshaft »

yeah... what Sprocc said... right on!!!

uh... is that "Mr. Sprocc" with the pointy ears and the tricorder?
/Greyshaft
Sprocc
Posts: 3
Joined: Sun Oct 02, 2005 11:20 pm

RE: Artifical Intelligence for World in Flames

Post by Sprocc »

No.

Sprocc is the character in a childrens book by Graham Base. Its one of my kids fav books
If I knew what I was doing, it wouldn't be called research.
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Sprocc

No.

Sprocc is the character in a childrens book by Graham Base. Its one of my kids fav books

Greyshaft: Consider this new knowledge part of your continuing education now that you have become a parent.
Steve

Perfection is an elusive goal.
User avatar
Greyshaft
Posts: 1979
Joined: Mon Oct 27, 2003 1:59 am
Location: Sydney, Australia

RE: Artifical Intelligence for World in Flames

Post by Greyshaft »

ORIGINAL: Shannon V. OKeets
Greyshaft: Consider this new knowledge part of your continuing education now that you have become a parent.

I'm certainly learning a lot!!! I could create a whole new color chart based on what I've found in Adam's diapers.
/Greyshaft
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Greyshaft

ORIGINAL: Shannon V. OKeets
Greyshaft: Consider this new knowledge part of your continuing education now that you have become a parent.

I'm certainly learning a lot!!! I could create a whole new color chart based on what I've found in Adam's diapers.
More information than I needed to know.
Steve

Perfection is an elusive goal.
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

I have been hammering away at designing the AI Opponent so it can play a respectable game (against you!). There is a ton of work to be done to handle all the little decisions that have to be made when playing WIF but I thought you might be willing to help me out at the highest level of the AIO’s design. What I am looking for are grand strategic plans for the AIO.

As I see it, China and France have limited strategic options and Italy has only a few choices because of its geographical position and smaller production and military capabilities. The other 5 major powers, on the other hand, have a lot of choices at the strategic level. I would like to have several strategic plans for each of the big 5 from which the AIO can choose, with a lot of optional little pieces that could be included or excluded. What I have created for the AIO in writing my other war games is an ambiguous starting position and early action choices/movement of units, so the human opponent isn’t sure exactly what the AIO intends to do next.

If you would like to provide me with some advice on strategic plans for any of the major powers, I would greatly appreciate it. If several people choose the same country, that would be fine too, because, after all, I am looking for multiple strategic plans. There are 11 scenarios in MWIF, so if you want to just advise on Barbarossa or one of the other scenarios, that would be great. The obvious starting place is the player notes are the end of the rule book, but here are some specifics that I am looking for:

(1) Victory cities to be taken and/or defended.
(2) Which major powers to declare war on, when, and any associated conditions concerning same.
(3) Which minor countries to declare war on, when, and any associated conditions concerning same.
(4) Which minor countries to align, when, and any associated conditions concerning same.
(5) Expected areas of conflict: with whom, where, type of combat (land, naval, air, convoys).
(6) A master production plan by unit type and/or gearing limits. This does not have to be detailed, just a broad outline.
(7) A time line for the strategic plan’s major milestones.

I have probably forgotten something, so add anything that you believe to be important.

I thank you for your continuing support.

Steve

P.S. I am still looking for people to help on the more detailed aspects of the AIO (Steve@PatternDiscovery.us)
Steve

Perfection is an elusive goal.
User avatar
Greyshaft
Posts: 1979
Joined: Mon Oct 27, 2003 1:59 am
Location: Sydney, Australia

RE: Artifical Intelligence for World in Flames

Post by Greyshaft »

I'll have a go for Italy...happy to share the topic if someone wanted to be Il Duce.
/Greyshaft
User avatar
Froonp
Posts: 7998
Joined: Tue Oct 21, 2003 8:23 pm
Location: Marseilles, France
Contact:

RE: Artifical Intelligence for World in Flames

Post by Froonp »

I intend to have a try at each major power one moment or the other.
Maybe for the ideas to be organized we should create a thread for each major power so that information is easy to collate ?
A thread named "Artifical Intelligence for World in Flames - Italy" for instance for Italy, and then anothere one for each major power.
What do you think ?
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Froonp
I intend to have a try at each major power one moment or the other.
Maybe for the ideas to be organized we should create a thread for each major power so that information is easy to collate ?
A thread named "Artifical Intelligence for World in Flames - Italy" for instance for Italy, and then anothere one for each major power.
What do you think ?

That's fine by me. However, let's only create the threads when someone has something to post to it (?).
Steve

Perfection is an elusive goal.
CBoehm
Posts: 113
Joined: Mon Oct 31, 2005 10:53 am
Location: Aarhus, Denmark

RE: Artifical Intelligence for World in Flames

Post by CBoehm »

If you would like to provide me with some advice on strategic plans for any of the major powers, I would greatly appreciate it.

Im willing to give a go at the CW which happens to be my favorite country exactly because of its multitude of strategic options. - should I post it here or email it to you?

A few detail questions?!
- you state that there will be no "master allied ai" ...does that mean that conditional strategies cannot be designed for the CW, ei. do this if US do that?
- "which VPhexes to garrison" how about other important cities/resource hexes?! (a key hexes are not VPhexes).
- can conditional ai behavior be implemented depending on which optional rules are used?
WIF the most wonderful, frustrating, uplifting and depressing of all games...
User avatar
Froonp
Posts: 7998
Joined: Tue Oct 21, 2003 8:23 pm
Location: Marseilles, France
Contact:

RE: Artifical Intelligence for World in Flames

Post by Froonp »

Im willing to give a go at the CW which happens to be my favorite country exactly because of its multitude of strategic options. - should I post it here or email it to you?
Please, open a new Thread named Artifical Intelligence for World in Flames - Commonwealth (or AI for MWiF - CW).
Thanks
Shannon V. OKeets
Posts: 22165
Joined: Wed May 18, 2005 11:51 pm
Location: Honolulu, Hawaii
Contact:

RE: Artifical Intelligence for World in Flames

Post by Shannon V. OKeets »

ORIGINAL: Froonp
Im willing to give a go at the CW which happens to be my favorite country exactly because of its multitude of strategic options. - should I post it here or email it to you?
Please, open a new Thread named Artifical Intelligence for World in Flames - Commonwealth (or AI for MWiF - CW).
Thanks

Done.
Steve

Perfection is an elusive goal.
Post Reply

Return to “AI Opponent Discussion”