1
DF Suggestions / Re: Possibility of AI integration in DF?
« on: April 16, 2023, 05:15:49 pm »
There's lots of types of AI for games, but based on your example and context I assume you mean specifically a deep-learning chat bot. There are some huge barriers.
Data-set: The reason Chat-GPT talks so realistically is because it has a massive database of transcribed dialogue, novels, and articles to learn the patterns of. If you want to make an immersive Dwarf Fortress chat bot, you would need a separate Dwarf Fortress data-set for every part of generation except grammar. Just creating enough of that data alone is a head-ache, So Tarn gotta write a couple Dwarf Fortress novels at minimum. This is made extra difficult by Dwarf Fortress having generative per-world lore, instead of game lore, so you have to set up a sorta Mad-Libs style system for history, places, and things.
Integration: Even if you get a DF Chat Bot perfectly staying in character and in world, you need a separate AI interpreter to actually make things happen according to the chat. For example, if you only had a chat bot and a representative of the humans said "We are declaring war on your civilization!" no war would happen. So you need to train a separate AI to translate text back into gameplay.
Offline: Chat bots are not nearly as optimized as txt-to-image diffusion yet, and you need a very strong computer to run it. DF is also a CPU heavy game. So the only real solution is to stream the chat bot. Even if you could easily deal with the logistics of running that server, it would force anyone using the feature to play online which would inconvenience everyone.
That said, there are lots of ways AI can be improved in DF that do not involve machine learning models. Invader AI can be improved by trying to break into doors, setting up their own traps outside, and having various battle formations / siege weapons. We can add diplomacy meetings with the classic multiple choice scripted style. Let's not get over our head and just focus on perfecting AI in the realm of programming before we jump to machine learning.
Data-set: The reason Chat-GPT talks so realistically is because it has a massive database of transcribed dialogue, novels, and articles to learn the patterns of. If you want to make an immersive Dwarf Fortress chat bot, you would need a separate Dwarf Fortress data-set for every part of generation except grammar. Just creating enough of that data alone is a head-ache, So Tarn gotta write a couple Dwarf Fortress novels at minimum. This is made extra difficult by Dwarf Fortress having generative per-world lore, instead of game lore, so you have to set up a sorta Mad-Libs style system for history, places, and things.
Integration: Even if you get a DF Chat Bot perfectly staying in character and in world, you need a separate AI interpreter to actually make things happen according to the chat. For example, if you only had a chat bot and a representative of the humans said "We are declaring war on your civilization!" no war would happen. So you need to train a separate AI to translate text back into gameplay.
Offline: Chat bots are not nearly as optimized as txt-to-image diffusion yet, and you need a very strong computer to run it. DF is also a CPU heavy game. So the only real solution is to stream the chat bot. Even if you could easily deal with the logistics of running that server, it would force anyone using the feature to play online which would inconvenience everyone.
That said, there are lots of ways AI can be improved in DF that do not involve machine learning models. Invader AI can be improved by trying to break into doors, setting up their own traps outside, and having various battle formations / siege weapons. We can add diplomacy meetings with the classic multiple choice scripted style. Let's not get over our head and just focus on perfecting AI in the realm of programming before we jump to machine learning.