The idea of universal intelligence is not a bug, it is a feature. It is mainly due to Legg/Hutter that we have that concept in the first place—and it is a fine one.
Not really. If you claim that a) intelligence is useful, and b) a maximally intelligent being that you have invented is useless … you made a mistake somewhere.
And their work is just the formalization of Solomonoff induction—the difficulty is in the derivation. People knew in advance that you can find the shortest theory to fit the data by taking a language, and then iterating up from the shortest expressible program until you find one that matches the data—it’s just that it’s not computable, which for now, means useless, and the exponential approximation isn’t much better.
Can you identify any working, useful system based on AIXI?
Okay, I don’t have a reference for them admitting that AIXI’s useless—but they acknowledge it’s uncomputable, and don’t have working code implementing it for an actual problem in a way better than existing “not intelligent” methods.
Solomonoff induction is concerned with sequence prediction—not decision theory. It is not a trivial extra step.
AIXI is also primarily concerned with sequence prediction and not decision theory.
“AIXI is a universal theory of sequential decision making akin to Solomonoff’s celebrated universal theory of induction. Solomonoff derived an optimal way of predicting future data, given previous observations, provided the data is sampled from a computable probability distribution. AIXI extends this approach to an optimal decision making agent embedded in an unknown environment.”
Right—but they know that. AIXI is a self-confessed abstract model.
IMO, AIXI does have some marketing issues. For instance:
“The book also presents a preliminary computable AI theory. We construct an algorithm AIXItl, which is superior to any other time t and space l bounded agent.”
That seems to be an inaccurate description, to me.
The idea of universal intelligence is not a bug, it is a feature. It is mainly due to Legg/Hutter that we have that concept in the first place—and it is a fine one.
Not really. If you claim that a) intelligence is useful, and b) a maximally intelligent being that you have invented is useless … you made a mistake somewhere.
And their work is just the formalization of Solomonoff induction—the difficulty is in the derivation. People knew in advance that you can find the shortest theory to fit the data by taking a language, and then iterating up from the shortest expressible program until you find one that matches the data—it’s just that it’s not computable, which for now, means useless, and the exponential approximation isn’t much better.
Can you identify any working, useful system based on AIXI?
I don’t think you have a reference for b).
Solomonoff induction is concerned with sequence prediction—not decision theory. It is not a trivial extra step.
Okay, I don’t have a reference for them admitting that AIXI’s useless—but they acknowledge it’s uncomputable, and don’t have working code implementing it for an actual problem in a way better than existing “not intelligent” methods.
AIXI is also primarily concerned with sequence prediction and not decision theory.
“AIXI is a universal theory of sequential decision making akin to Solomonoff’s celebrated universal theory of induction. Solomonoff derived an optimal way of predicting future data, given previous observations, provided the data is sampled from a computable probability distribution. AIXI extends this approach to an optimal decision making agent embedded in an unknown environment.”
http://www.hutter1.net/ai/uaibook.htm
Okay, you’re right, my apologies. The point about uncomputability and uselessness of the decision theory still stands.
Right—but they know that. AIXI is a self-confessed abstract model.
IMO, AIXI does have some marketing issues. For instance:
“The book also presents a preliminary computable AI theory. We construct an algorithm AIXItl, which is superior to any other time t and space l bounded agent.”
That seems to be an inaccurate description, to me.