coregames
Posts: 470
Joined: 8/12/2004 Status: offline
|
quote:
ORIGINAL: Shannon V. OKeets There was a reference to the AI software for chess programs and its decision tree approach. In case you aren't aware, WiF is a much more complex game that chess. There are more pieces. There are more unit types. The playing surface is larger. Instead of moving one piece per turn you move dozens. Combat is probabilistic not deterministic. The terrain is diverse and changes with the season. Need I go on? If you can't assemble a grand strategy in WiF, you can't succeed. Obviously, an AI can't create a decision tree for every land, air and naval move possible for turns (or even impulses) in advance. My point was not to micro-control tactics with a decisions tree (chess was only an analogy). But something must connect strategic themes into a cohesive whole, or the AI's position will drift aimlessly into ruin. It is true that WiF uses a lot of luck, but if that was the biggest determining factor, I for one would not play it. There are enough instances of luck for the effects of it to be somewhat statistically averaged out (especially if you play many times). quote:
Operational plans would require the development of a special language. The language, at a minimum would list objectives, a time table, a maximum level of troops commitment. Am I to assume that the "special language" you propose for operational plans is one-dimensional, with no branching allowed? If branching does exist in the way the operational plans are formulated (as in having a Plan B), that sounds a lot like a decision tree to me (though admittedly not a tactical one). If no branching exists, in my opinion the AI will be weak, following a shortsighted and inflexible game plan. In the middle of a long scenario, operations are sometimes planned 6 turns or more in advance, and the game is so variable that standardized strategies past a certain point are of limited value. I am not an advanced programmer, and I don't know what goes into a strong AI for a complex grand strategic computer wargame. I have played a lot of WiF, however. To model what strong WiF players do strategically, the AI will need to use some sort of lookahead, weighing various possible outcomes against eachother based on some kind of valuation of the position. If you read my post, you will notice that my point was not about using brute-force lookahead, but about positional recognition: quote:
Strong chess programs are increasingly dependent on positional parameters, still relying on brute-force look-ahead to see many variations to a great depth, but having a better idea of how to evaluate the position at the termination of each branch of their search. To play well, MWiF will need to have a strong meta-strategy component... I go on to talk about using a high-level meta-strategy "tree" that is constructed by part of the AI, tasked with stringing together these "called plays" from a library of plans that represent the experience of strong players (Greyshaft mentions this in an earlier post in this thread). The result is not an attempt to use brute-force to plan an impulse or a turn, but rather, to prepare well in advance so that the most important operations have an improved chance of success. I do understand that WiF is much more complex than chess, and that it is unpredictable. Everything I say may be way off-base, given my lack of experience developing AI. I am willing to be shown why a meta-strategic decision tree is a bad idea.
< Message edited by coregames -- 5/20/2005 7:27:37 AM >
|