I like planning for a reason: non hierarchical planning is the most powerful and easiest to implement algorithm for a simple AI. Yes, it has some problems, I can't remember the specific context on which states should be added to that of the global state (so the global state isn't really a state).
AI approaches discussion (split from Interesting AI Examples thread)
non hierarchical planning is the most powerful and easiest to implement algorithm for a simple AI.
Highly debatable.
Yes, it has some problems, I can't remember the specific context on which states should be added to that of the global state (so the global state isn't really a state).
This gives me the impression you're not nearly familiar enough with the technique, nor are you really qualified to be recommending it above all other approaches.
[edit] You're also off-topic as this has nothing to do with the OP's question.
Author
Above all other more difficult to implement techniques? You really have a hard case to make, APOCH, really. It takes not that many lines to write a basic planning algorithm, I can only see a hardcoded logic being easier but not as powerful.
I like planning for a reason: non hierarchical planning is the most powerful and easiest to implement algorithm for a simple AI. Yes, it has some problems, I can't remember the specific context on which states should be added to that of the global state (so the global state isn't really a state).
Wot?!?!
Above all other more difficult to implement techniques? You really have a hard case to make, APOCH, really. It takes not that many lines to write a basic planning algorithm, I can only see a hardcoded logic being easier but not as powerful.
Writing the algorithm is not the same as teasing out good behavior. By that measure, picking random behaviors would be the best possible approach because it is easy to implement.
Author
Context of work? Yes, I agree with that.
Talking about interesting AI and lazy/easy (I really favor easy), it's possible to automatically adjust weights if you find a way to judge the AI, maybe like making a quadruped gait (
) in a more efficient way that a human could do... but that's just something useless and a proof of concept I made. The thing is that we've seen stuff like this in the field, this is tested stuff that works: there're cases in which you can just spend time extracting and formating important data from the game.
Your first post speaks of planners, but now you're talking about machine learning. What are you actually advocating for, exactly?
I've worked on game AI for over 15 years. I've never seen a planner used in production. That's not to say that they are bad tools, or that nobody uses planning; quite the contrary. It's just that there are other tools which are superior in their respective areas.
Machine learning is even less appealing for games, for well-established reasons that I won't retread unless you for whatever reason can't/won't dig up any of those debates yourself.
I've worked on game AI for over 15 years. I've never seen a planner used in production. That's not to say that they are bad tools, or that nobody uses planning; quite the contrary. It's just that there are other tools which are superior in their respective areas.
Machine learning is even less appealing for games, for well-established reasons that I won't retread unless you for whatever reason can't/won't dig up any of those debates yourself.
Author
I've a trait: When I talk I tend to go all over the place (I really mean it). That second part of my post wasn't really directed what I previously said, so I got a warning out of nothing :/ .
Yes, heuristics (which sounds less buzzwordy than ML) as an optimization technique. Simulated Annealing which did a a decent job at go (in the 2000s). I think they implemented over decition space (which in go overlaps a lot)... GAs could also be used, but i've not heard of that. There're statistical based methodologies too, but I've yet to dig into that. There're niches, thats all, I'm sure these things could take over (when a fitness/evaluation function could be available).
I really do not want to go through my "and this is why any sort of learning in game AI sucks all to hell" speech. I'm going to have to repeat it enough times at GDC in 2 weeks. *sigh*
I really do not want to go through my "and this is why any sort of learning in game AI sucks all to hell" speech. I'm going to have to repeat it enough times at GDC in 2 weeks. *sigh*
I'll state the big one: The behavior of NPCs in games should be optimized for things like how much fun the player has, or how engaged she is, or how convincing is the illusion that the agents are real. These are measures that are very hard to work with in machine learning.
For games where the measure of quality is somewhat objective (chess, go, etc.), machine learning is a reasonable approach; but that's not the kind of game we are usually concerned with in this forum.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
