Interview on artificial intelligence with the Gazette du Wargamer

Gazette du Wargamer interviewed VR Designs a few weeks ago on artificial intelligence design in wargames. Here is the english version of the interview:

1/ For you, what’s really an AI in a wargame. Several scripts and algorithms, of course, but anything more ?

V.R. For me it is most difficult and most challenging part of coding a wargame. It is really really difficult to make one. And its even harder to be fully satisfied with one. Every AI I make always ends up as a never-ending project. You could spend multiple lifetimes on making a good wargame AI. Really!

2/ For a PC wargame, what are the differents AI models ? And the differents software tools for developers ?

V.R. I don’t know. I code every AI I make from scratch. I don’t copy-cat any algorithms and do not use 3rd party libraries.

However I do see basically 2 types of AI you can code for a wargame. First there is the chess-based approach in which I let the AI iterate through all the moves without any preconception and for example use a scoring mechanism to judge the likely outcomes of those moves and decide how desirable those outcomes are. Second is the fuzzy-based approach in which I use fixed rules to determine the AI moves. The first approach is really slow and calculation heavy and the second approach is really fast. Sometimes I mix these approaches a little bit.

3/ What is possible with the computing power of a modern PC (more or less 2 years old) of good quality ? 3D asides, what other technicals constraints is there ?

V.R. Well no matter how fast a good modern PC is. You run out of time pretty quickly if you have a game of 400+ units and each unit can chose between 100s of moves. I am not sure how familiar readers are with permutation mathematics but even all PCs on the whole earth could probably not handle checking all possible combinations of moves for the opening round. Let’s not ever consider running some calculations a few rounds in to the future.

4/ Are some kind of wargames, so of historical conflicts, or scales of games, more appropriate for (current) AI ? If so, which one ?

V.R. Yes. Games with a very low unit count. Period.

5/ From your experience in game design, what technicals challenges have you meet ? Which one did you overcome ? Which one are still a barrier ?

V.R. I learned a lot coding different AIs for Advanced Tactics, Warsaw To Paris, Case Blue and Small General. The key thing I learned is that when I ever start a new engine I’ll have to adjust the design of the game to the AI instead of the other way around.

6/ Is an AI allied with the player (for a game including alliances and several factions) more simple or more complicated to build than a unique AI enemy of the player (for a game where there’s mainly one adversary) ?

V.R. I actually avoided coding this option, since to do this in a meaningfull way is really difficult. Team play with an AI requires some form of communication with the AI. You can see the problem right? For example: How to agree with the AI who takes which part of the frontline? How to divide the resources? How to coordinate doing encirclements or retreats with mixed up units ?

7/ In strategy games and wargames AI are very often better in attack than in defense, why ?

V.R. I am not sure that’s true. I found coding a defensive AI is actually easier. Though I might want to make an exception for a retreating AI, my AIs are usually pretty good in standing and fighting, but to retreat at the right moment is a hard skill to teach, as is reforming a shattered or damaged front somewhere further behind the lines.

8/ To what extent the perception of the AI from the player has to be taken into account ? Can an AI in a game just be less good strategically (than a human adversary), as long as it offers a good gameplay with its actions ? And so at the end, for a game, offers an entertaining result, thus something satisfying ?

V.R. It might be impossible to create an AI that is a von Manstein… But it should be possible to create an AI that is a Paulus or a Gamelin. I think mimicking the historical moves of these commanders is more than feasible with current AIs. However the limits of AI design usually show if the AI is playing the side that has to perform a feat of genius and mobility like for example the German offensive of the 3rd battle for Kharkov or Rommel in Africa. Also the limits of AI design show if you make games that are playable against Human players as well as AIs. The good human player will always do better, and the player will than often feel the AI is inadequate.

9/ For wargames, what development can we expect for AI in a relatively near future ? A crazy idea probably, if that was possible technically, could a wargame having part of its AI in the future (and rather hypothetical) « cloud » gain an advantage (ex : computing power, comparing moves in same other games) ?

V.R. Yes sure I think that’s possible. However raw calculation power can do only limited wonders. Talking about clouds, I myself toyed around with the idea of designing a wargame that uploads the AI experiences back to a central server and shares them with all the other AIs of all the other game copies in circulation. This way the AI might actually gather a database of what are likely human moves and reactions. The only thing that’s holding me back is that this sort of stuff is not like exactly easy coding

10/ A chess game has a board with 64 squares and 32 pieces. So a base less important than a good wargame, and far much less than some « monster wargames ». With 256 processors in parallel (allowing 200 millions move / sec.), in 1996-97 Deep Blue has finally beaten the world chess champion Gary Kasparov.

In theory, well programmed, a computer do not make mistakes. But even with a super-computer, the machine still lacks a strategical vision, a global view ? If we increase the size of the « board game », for a wargaming simulation is it just a question of computing power ?

V.R. I already said something about this in question 3. The thing is… it is not only the number of squares and the number of pieces. It’s mainly the fact that in chess you can move 1 piece per round, In a wargame you can move all pieces every round. No super-computer can handle a rough processing power based AI approach for any serious wargame.

11/ Tiebreaker question : If in a wargame an AI could beat certainly an experienced wargamer, whatever the simulated theatre of operations, does a good game have to limit the AI to necessarily let the player win ?

V.R. This is a very speculative question. As I said earlier: I do not see this happening. But well… if it happens… Yes then you would have to code in some low level difficulty settings in which you actually handicap the AI. Let it misjudge situations… Let it forget to see certain units… It should be fun to code

Note: Keep an eye on Gazette du Wargamer for interviews on AI design with other wargame designers.

This entry was posted in Artificial Intelligence, DC:Case Blue, Game Design. Bookmark the permalink.

Leave a Reply