AIXI is highly incompetent (or even ill-defined) in the continuous domain. The world does not come ready-divided into episodes. Actions do not come in discrete chunks. AIXI has no mechanism for creating abstract categories, such as a "heavy object" or "animal".
AIXI suffers from the same problems seen in Reinforcement Learning in general. After you have eaten a large meal, you gain no reward for eating even more soon afterwards. In the real world, reward is attached to satiating urges.
@otonanoC When you say that the world cannot be divided into discrete parts, exactly how do you define "discrete"? For one thing, Planck's constant and Planck lengths pretty much define the world as discrete chunks. Not an absolute division, as in point T is Planck X, but a limit for observation and physical effects. Humans, the best (unfortunately) example we have of rational agents, deal with time on a very limited basis (ms reaction time).
@otonanoC Both the lack of abstract categories and generality of rewards are left to the definition of a problem. You could easily add as complex a reward function as you like, and define as large a set of categories, actors, and actions, as you so desired.
Of course, in the 'real world', an agent has to define its own problem and build its own knowledge. What would be the primary goal (reward function?) of such a "practical" intelligent agent?
Good to see you here. Allow me to continue my attack on AIXI. One thing none of the researchers emphasize, which should be emphasized after all is this: The agent is only maximizing EXPECTED REWARD. There is no gaurantee that the agent will actually be optimal in practice. The distinction I am making is subtle but extremely important. Maximizing expected reward may not entail an actual maximization of reward in reality. Please see me next paragraph as I am running out of
How then, are actual biological animals so successful in practice? The answer is that animals have instinctual behaviors and dispositions that are bequeathed to them by natural selection. This includes human beings. I predict that in the coming years, AGI's main problem will be how to unite instinctual behaviors that work beautifully in practice together with their tabula-rasa inductive prediction paradigm.
@otonanoC Well, the instinctive behavior is just a subset of reward goals -- instincts are essentially the most prioritized reward goals; it's just that nature's evolutionary schemes have hand chosen the most important goals and integrated them more thoroughly. For all such goals, there must be some kind of reward function which predicts how well that goal is met (for instance, hunger -- is this edible, how nutritious is it, how much of it can I get, etc.).
@otonanoC Most animals lack much intelligence, and instead only have the instincts necessary for survival (in a specific niche)... Whereas, for an AI, simple survival is pointless -- an infinite loop could accomplish that! You could argue each animal is fine-tuned for its niche, but that's just weak AI in practice. Humans are the closest (but not perfect) example to strong AI that we have.
How can you say in one breath that human beings are the closest example to strong AI, and in the next breath ignore the fact that human beings are embodied animals in a narrow context of the physical world of the earth? Do you believe intelligence exists in the universe for no reason related to their environment in which it is found?
This clever trick where we redefine all the words we are using ahead of time in order to ensure that our mathematical proofs will work out the way we intended -- this is a game I do not intend to play with you.
@otonanoC So I might argue that rather than take a look at some kind of innovative architecture for the mind of an AI, we instead look at better ways to design two things we know it needs to have. AIs can't learn new actions, unlike humans, which learn new skills without necessarily even being taught. They also can't see their own reward function's inefficiencies, and so must have a meta-reward function. How to design a meta-reward function and emergent behavior is the hard part.
Why are putting "real world" and "practical" in scare quotes here? I am not and I have not argued with the mathematical proofs and mathematical arguments of AIXI. I find this very peculiar that you always expect there to be a human in the loop of this AIXI agent who gives the "definition of a problem". You have ignored the very salient points I have made about an agent in a real setting.
In saying **you could easily add** -- what you really meant to say there is that the mathematical argument of AIXI **is unaffected by** the addition of these complex reward functions.
The easiness or uneasiness of adding such things in an effective embodied agent is a completely different issue. So you did not really mean "easy" there. What you meant is that the argument is unaffected by these things.
So that there is no confusion between the two of us let me emphasize this again --> I am not in any shape or form claiming I have found a logical error in the proof of AIXI as a general RL agent. I would hope your replies would stop putting that in my mouth.
Having said that, it is the lingering, often unspoken claim that such an agent will be effective in practice that is the source of all my arguments in these comment boxes.
@otonanoC If your point that is not practical to implement such an AIXI in real world applications, then that is only a result of human failure to do so. Any sufficiently advanced set of reward functions and knowledge representations should be capable of managing its own measure of success. Your argument, if I am reading correctly, is that instinctual behavior and adaptation to niches are essential. This is true, but is essentially the same as prioritized reward functions and genetic algorithms.
Except those two things are not essentially the same at all! In the real world there are problems of morphology. Certainly no one is going make the claim that your embodied AIXI agent will have the capacity to rebuild its own body continually. In the real world there are time restraints, limits on space, and limited resources. We should not confuse the mathematical optimality of these equations with successful action in the real physical world.
@otonanoC And constraints on time, space, and other resources are somehow mathematically unexplainable? There are entire fields of study devoted to representing such problems in the most mathematically optimal way, and they do in fact promote successful actions; it's how people manage businesses and governments. How are instinctual behaviors and adaptations somehow outside mathematical (or algorithmic) description?
I never said anything about this being outside of mathematical description. (putting words in my mouth yet again!) I think you and I need to take up this conversation in private messages because I'm sick and tired of the character limit on these stupid comment boxes. I assure you I can explain all of this very clear to you in private messages.
Just for now, understand that you should not confuse the maximization of Expected Reward with the actual accumulation of actual Reward in practice during a run. They are not the same thing, and that is because the environment is partially observed.
Even if YOU DID have maximum reward accumulation, you should not confuse that with actual effective performance in practice. And the defense of that statement cannot fit in these comment boxes! voila!
AIXI is highly incompetent (or even ill-defined) in the continuous domain. The world does not come ready-divided into episodes. Actions do not come in discrete chunks. AIXI has no mechanism for creating abstract categories, such as a "heavy object" or "animal".
AIXI suffers from the same problems seen in Reinforcement Learning in general. After you have eaten a large meal, you gain no reward for eating even more soon afterwards. In the real world, reward is attached to satiating urges.
otonanoC 4 months ago
For a model of intelligence that is truly outside the paradigm of RL, I recommend the work of Joscha Bach, or that of Rolf Pfeifer.
otonanoC 4 months ago
@otonanoC When you say that the world cannot be divided into discrete parts, exactly how do you define "discrete"? For one thing, Planck's constant and Planck lengths pretty much define the world as discrete chunks. Not an absolute division, as in point T is Planck X, but a limit for observation and physical effects. Humans, the best (unfortunately) example we have of rational agents, deal with time on a very limited basis (ms reaction time).
GeekProdigyGuy 3 months ago
@otonanoC Both the lack of abstract categories and generality of rewards are left to the definition of a problem. You could easily add as complex a reward function as you like, and define as large a set of categories, actors, and actions, as you so desired.
Of course, in the 'real world', an agent has to define its own problem and build its own knowledge. What would be the primary goal (reward function?) of such a "practical" intelligent agent?
GeekProdigyGuy 3 months ago
@GeekProdigyGuy
Good to see you here. Allow me to continue my attack on AIXI. One thing none of the researchers emphasize, which should be emphasized after all is this: The agent is only maximizing EXPECTED REWARD. There is no gaurantee that the agent will actually be optimal in practice. The distinction I am making is subtle but extremely important. Maximizing expected reward may not entail an actual maximization of reward in reality. Please see me next paragraph as I am running out of
otonanoC 3 months ago
How then, are actual biological animals so successful in practice? The answer is that animals have instinctual behaviors and dispositions that are bequeathed to them by natural selection. This includes human beings. I predict that in the coming years, AGI's main problem will be how to unite instinctual behaviors that work beautifully in practice together with their tabula-rasa inductive prediction paradigm.
otonanoC 3 months ago
@otonanoC Well, the instinctive behavior is just a subset of reward goals -- instincts are essentially the most prioritized reward goals; it's just that nature's evolutionary schemes have hand chosen the most important goals and integrated them more thoroughly. For all such goals, there must be some kind of reward function which predicts how well that goal is met (for instance, hunger -- is this edible, how nutritious is it, how much of it can I get, etc.).
GeekProdigyGuy 3 months ago
@otonanoC Most animals lack much intelligence, and instead only have the instincts necessary for survival (in a specific niche)... Whereas, for an AI, simple survival is pointless -- an infinite loop could accomplish that! You could argue each animal is fine-tuned for its niche, but that's just weak AI in practice. Humans are the closest (but not perfect) example to strong AI that we have.
GeekProdigyGuy 3 months ago
@GeekProdigyGuy
How can you say in one breath that human beings are the closest example to strong AI, and in the next breath ignore the fact that human beings are embodied animals in a narrow context of the physical world of the earth? Do you believe intelligence exists in the universe for no reason related to their environment in which it is found?
otonanoC 2 months ago
@GeekProdigyGuy
This clever trick where we redefine all the words we are using ahead of time in order to ensure that our mathematical proofs will work out the way we intended -- this is a game I do not intend to play with you.
otonanoC 2 months ago
@otonanoC So I might argue that rather than take a look at some kind of innovative architecture for the mind of an AI, we instead look at better ways to design two things we know it needs to have. AIs can't learn new actions, unlike humans, which learn new skills without necessarily even being taught. They also can't see their own reward function's inefficiencies, and so must have a meta-reward function. How to design a meta-reward function and emergent behavior is the hard part.
GeekProdigyGuy 3 months ago
@GeekProdigyGuy
Why are putting "real world" and "practical" in scare quotes here? I am not and I have not argued with the mathematical proofs and mathematical arguments of AIXI. I find this very peculiar that you always expect there to be a human in the loop of this AIXI agent who gives the "definition of a problem". You have ignored the very salient points I have made about an agent in a real setting.
otonanoC 2 months ago
@otonanoC
In saying **you could easily add** -- what you really meant to say there is that the mathematical argument of AIXI **is unaffected by** the addition of these complex reward functions.
The easiness or uneasiness of adding such things in an effective embodied agent is a completely different issue. So you did not really mean "easy" there. What you meant is that the argument is unaffected by these things.
otonanoC 2 months ago
@GeekProdigyGuy
So that there is no confusion between the two of us let me emphasize this again --> I am not in any shape or form claiming I have found a logical error in the proof of AIXI as a general RL agent. I would hope your replies would stop putting that in my mouth.
Having said that, it is the lingering, often unspoken claim that such an agent will be effective in practice that is the source of all my arguments in these comment boxes.
otonanoC 2 months ago
@otonanoC If your point that is not practical to implement such an AIXI in real world applications, then that is only a result of human failure to do so. Any sufficiently advanced set of reward functions and knowledge representations should be capable of managing its own measure of success. Your argument, if I am reading correctly, is that instinctual behavior and adaptation to niches are essential. This is true, but is essentially the same as prioritized reward functions and genetic algorithms.
GeekProdigyGuy 2 months ago
@GeekProdigyGuy
Except those two things are not essentially the same at all! In the real world there are problems of morphology. Certainly no one is going make the claim that your embodied AIXI agent will have the capacity to rebuild its own body continually. In the real world there are time restraints, limits on space, and limited resources. We should not confuse the mathematical optimality of these equations with successful action in the real physical world.
otonanoC 2 months ago
@otonanoC And constraints on time, space, and other resources are somehow mathematically unexplainable? There are entire fields of study devoted to representing such problems in the most mathematically optimal way, and they do in fact promote successful actions; it's how people manage businesses and governments. How are instinctual behaviors and adaptations somehow outside mathematical (or algorithmic) description?
GeekProdigyGuy 2 months ago
@GeekProdigyGuy
I never said anything about this being outside of mathematical description. (putting words in my mouth yet again!) I think you and I need to take up this conversation in private messages because I'm sick and tired of the character limit on these stupid comment boxes. I assure you I can explain all of this very clear to you in private messages.
otonanoC 1 month ago
@GeekProdigyGuy
Just for now, understand that you should not confuse the maximization of Expected Reward with the actual accumulation of actual Reward in practice during a run. They are not the same thing, and that is because the environment is partially observed.
Even if YOU DID have maximum reward accumulation, you should not confuse that with actual effective performance in practice. And the defense of that statement cannot fit in these comment boxes! voila!
otonanoC 1 month ago
Can you please pump up your recording volume, attempt automatic subtitling with your script, and record some more spontaneous videos?
skdkskdk 1 year ago