![]() ![]() Neural networks were originally devised as function approximators for classification problems. diplomacy models.Īnd while I could write about other forms of AI, I've been leading this up to neural networks so let's talk about those. But again, I feel like perhaps for certain sub-problems in Stellaris there could be an application. So, could this be done? Probably not for Stellaris as a whole. In addition, you need to model the Stellaris game state in some way, and the Stellaris game state is ludicrously complex. ![]() This is extremely non-trivial for hard problems. The issue is learning or developing the transition model. So how does this help? Well, you're still using search, but you're approximating a lot of the dimensions to the search. A particular action applied to a certain state will result in a new state (the transitions can be probablistic as well). This consist of a defined set of actions that can be applied to a game state space. (I'm getting to neural networks I swear). Is it hopeless to use search techniques? Probably not actually, but it is certainly non-trivial.Īnother form of AI that could be useful for game playing are state transition algorithms, e.g. It is difficult to guarantee performance on a search, and time limiting it may lead to very irregular and bizarre results especially in complex situations which is precisely when you want a good answer. So why don't they do this? To put it simply, performance. searching for what buildings to build next. I think it is obvious why search cannot be used broadly for a game like Stellaris however, there may be sub-problems to which it may apply. Basically, a move looks like it is good after the Nth move but at some point after the Nth move it actually is guaranteed to lead to a loss. The drawback to search is that the search spaces becomes very large very fast even for simple games (usually such spaces are referred to as intractable), additionally, unless you can guarantee to search to a winning state, the AI becomes subject to what is called an N-level trap. There are a lot of search algorithms, but in short, the idea is to construct a space that defines the legal moves from a given current game state and then search through the possible moves for a winning move, or a move that leads to the best possible position. The other major form of early attempts at AI for games (and other problems, more on this shortly) was searching. No matter how much the devs add or modify the rules, there are gaps and limitations. So you can see this in video game AI, like that used in Stellaris. Two, there are limitations to the extent of what an expert system can learn since it is confined to formally logical spaces, and again there are certain types of derived knowledge that are difficult to represent. One, it is very difficult to encapsulate some types of expert knowledge as logic. At one time, expert systems were believed to be the future of AI however, they've fallen out of favor for two reasons. In a sense, the developers provide this training step by iteratively adding to or revising the game AI's logic however, to my knowledge this is rarely done in a formal algorithmic fashion. ![]() A training algorithm is then run on the set of logical rules to train the system, in which it learns additional logical rules about the system. What is an expert system? An expert system is a set of logical representations of expert knowledge. In some very loose sense, you could liken most video game AIs to untrained expert systems. To begin, in the vast majority of video games, the AI does not use actual AI techniques. Note, I'm not per se an expert on video game artificial intelligence, but I have a fair amount of academic background, including a fair number of publications in high quality conferences and journals (note by any reasonable measure I'm an academic lightweight relative to the big names in the field, I do not purport to be of that caliber), in the application and theory of AI. P I'm more than happy to answer any specific questions. Some people with a high level ofexperience with AI will probably notice these "errors" but I'm not going to write something long, detailed and technical mainly because I don't want to as I'm currently on vacation. I'm going to keep this relatively short and simple, and I will also try to avoid being too technical. (N.B.: This is a copy and paste from another thread because I am *not* rewriting this every time there's a question on using actual AI techniques in Stellaris ) ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |