Ever more general problem-solvers
Dec. 17th, 2012 08:55 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
davvApparently some AIs can learn on their own without even being told the rules of the game they're playing, but just by being rewarded or penalized. Granted, this is only for toy problems, but it does formalize the problem and solution of reinforcement learning-based artificial intelligence.
Interestingly, if you look at it one way, the AI doesn't know its utility function. In another way, its utility function, what it wants to maximize, is simply to be rewarded the most by the giver of such rewards.
In here there is a story about "careful what you wish for". Still, I think a golem is more likely than a Terminator. And in either case, I don't think it's going to be based on the system above. Instead I think there's a balance: some intelligence functions are very fast but narrow (like A* optimal routing), and other are very general but very slow (like minimax). The thing above is closer to evolution (general and slow) than thought (specialized and fast). But I could be wrong, of course :) And even though minimax is slow in absolute terms, chess AIs using it still beat world-class players.
(And I also think that if further development of general-purpose AI occurs, it will pull us closer to a Singerian position: it doesn't matter as much whether a creature can think as whether it can feel (in the sense of intentionality). That in itself dissolves the problem of artificial intelligence by saying that "intelligence", problem-solving, is not actually the special matter, consciousness is. Perhaps consciousness informs intelligence, or works as some kind of "insight generator" that could also be made, if less effectively, by more mundane systems. It does not follow, though, that consciousness reduces to intelligence.)
((And it's late, and this is informal.))
Interestingly, if you look at it one way, the AI doesn't know its utility function. In another way, its utility function, what it wants to maximize, is simply to be rewarded the most by the giver of such rewards.
In here there is a story about "careful what you wish for". Still, I think a golem is more likely than a Terminator. And in either case, I don't think it's going to be based on the system above. Instead I think there's a balance: some intelligence functions are very fast but narrow (like A* optimal routing), and other are very general but very slow (like minimax). The thing above is closer to evolution (general and slow) than thought (specialized and fast). But I could be wrong, of course :) And even though minimax is slow in absolute terms, chess AIs using it still beat world-class players.
(And I also think that if further development of general-purpose AI occurs, it will pull us closer to a Singerian position: it doesn't matter as much whether a creature can think as whether it can feel (in the sense of intentionality). That in itself dissolves the problem of artificial intelligence by saying that "intelligence", problem-solving, is not actually the special matter, consciousness is. Perhaps consciousness informs intelligence, or works as some kind of "insight generator" that could also be made, if less effectively, by more mundane systems. It does not follow, though, that consciousness reduces to intelligence.)
((And it's late, and this is informal.))