I'm far from proficient in this space, but as I see it:
- The key problem in designing game AI lies in signalling to the human player why the AI acts the way it does; it's no fun to be beaten by a hypercompetent AI that's using a strategy you have no insight into. If you can't figure out what weakness of yours the AI exploited and learn to not make that mistake again, you might as well be playing Super Hexagon. Machine Learning approaches basically creates a blackbox where the internals are incomprehensible by humans even if the output makes sense, so there's an intrinsic challenge there in giving the player insight into how the AI makes decisions.
- GSGs and other strategy games have a huge advantage here in that decisions are less reaction time dependent than in e.g. FPS or racing games. On the other hand, the number of variables and possible decision space is much, much larger. Probably some sort of sub-system approach would have to be taken, where an AI designer decides which variables should be most relevant for a warfare AI, which are important for an economic AI, etc, in order to guide the training of the model(s) in the right direction.
- As mentioned above, the model's ability to remember what it has done before is really important, since "strategies" are by definition multi-step affairs executed over time. From what I gather most LLMs aren't well tuned to this, but I've been told the tech to "remember" decisions made in the same session does exist.
It's interesting and not too far-fetched, but my gut tells me it's probably going to be a while until it's operationalized.
- The key problem in designing game AI lies in signalling to the human player why the AI acts the way it does; it's no fun to be beaten by a hypercompetent AI that's using a strategy you have no insight into. If you can't figure out what weakness of yours the AI exploited and learn to not make that mistake again, you might as well be playing Super Hexagon. Machine Learning approaches basically creates a blackbox where the internals are incomprehensible by humans even if the output makes sense, so there's an intrinsic challenge there in giving the player insight into how the AI makes decisions.
- GSGs and other strategy games have a huge advantage here in that decisions are less reaction time dependent than in e.g. FPS or racing games. On the other hand, the number of variables and possible decision space is much, much larger. Probably some sort of sub-system approach would have to be taken, where an AI designer decides which variables should be most relevant for a warfare AI, which are important for an economic AI, etc, in order to guide the training of the model(s) in the right direction.
- As mentioned above, the model's ability to remember what it has done before is really important, since "strategies" are by definition multi-step affairs executed over time. From what I gather most LLMs aren't well tuned to this, but I've been told the tech to "remember" decisions made in the same session does exist.
It's interesting and not too far-fetched, but my gut tells me it's probably going to be a while until it's operationalized.
- 16
- 14
- 2
