There’s a pretty neat and general way of abstracting learning systems out into agent, ultimate values and environment using the framework of reinforcement learning.
Interesting. I’d appreciate references or links. To me, the interesting and still open question is how these “ultimate” values relate to the outcome of rational reflection and experimentation by the individual.
To me, the interesting and still open question is how these “ultimate” values relate to the outcome of rational reflection and experimentation by the individual
So: changes to ultimate values can potentially happen when there are various kinds of malfunction. Memetic hijacking illustrates one way in which it can happen. Nature normally attempts to build systems which are robust and resistant to this kind of change—but such changes can happen.
Maybe existing victims of memetic hijacking could use “reflection and experimentation” to help them to sort their heads out and recover from the attack on their values.
Thanks for the links. Both the AIXI and the Machine Super Intelligence use cardinal utilities, or in the latter case rational-number approximations to cardinal utilities (not sure if economists have a separate label for that), for their reward functions. I suspect this limits their applicability to human and other organisms.
Maybe existing victims of memetic hijacking could use “reflection and experimentation” to help them to sort their heads out and recover from the attack on their values.
In some cases. But the whole concept of “rationality” can probably usefully be viewed as a memeplex. And rational reflection leading to its rejection, while not a priori impossible, seems unlikely.
The good news from a gene’s point of view—in case anyone still cares about that—is that our genes probably co-evolved with rationality memes for a significant time period. Lately, though, the rate of evolution of the memes may be leaving the genes in the dust. That is, their time constants of adaptation to environmental change differ dramatically.
Both the AIXI and the Machine Super Intelligence use cardinal utilities, or in the latter case rational-number approximations to cardinal utilities (not sure if economists have a separate label for that), for their reward functions. I suspect this limits their applicability to human and other organisms.
FWIW, I don’t see that as much of a problem. I’m more concerned about humans having a multitude of pain sensors (multiple reward channels), and a big mountain of a-priori knowledge about which actions are associated with which types of pain—though that doesn’t exactly break the utility-based models either.
But the whole concept of “rationality” can probably usefully be viewed as a memeplex. And rational reflection leading to its rejection, while not a priori impossible, seems unlikely.
Sure, but “rationality” and “values” are pretty orthogonal ideas. You can use rational thinking to pursue practically any set of values. I suppose if your values are crazy ones, a dose of rationality might have an effect.
Lately, though, the rate of evolution of the memes may be leaving the genes in the dust.
Yes indeed. That’s been going on since the stone age, and it has left its mark on human nature.
Interesting. I’d appreciate references or links. To me, the interesting and still open question is how these “ultimate” values relate to the outcome of rational reflection and experimentation by the individual.
I just mean the cybernetic agent-environment framework with a reward/utility signal. For example, see page 1 of Hibbard’s recent paper, page 5 of UNIVERSAL ALGORITHMIC INTELLIGENCE A mathematical top!down approach, or page 39 of Machine Super Intelligence.
So: changes to ultimate values can potentially happen when there are various kinds of malfunction. Memetic hijacking illustrates one way in which it can happen. Nature normally attempts to build systems which are robust and resistant to this kind of change—but such changes can happen.
Maybe existing victims of memetic hijacking could use “reflection and experimentation” to help them to sort their heads out and recover from the attack on their values.
Thanks for the links. Both the AIXI and the Machine Super Intelligence use cardinal utilities, or in the latter case rational-number approximations to cardinal utilities (not sure if economists have a separate label for that), for their reward functions. I suspect this limits their applicability to human and other organisms.
In some cases. But the whole concept of “rationality” can probably usefully be viewed as a memeplex. And rational reflection leading to its rejection, while not a priori impossible, seems unlikely.
The good news from a gene’s point of view—in case anyone still cares about that—is that our genes probably co-evolved with rationality memes for a significant time period. Lately, though, the rate of evolution of the memes may be leaving the genes in the dust. That is, their time constants of adaptation to environmental change differ dramatically.
FWIW, I don’t see that as much of a problem. I’m more concerned about humans having a multitude of pain sensors (multiple reward channels), and a big mountain of a-priori knowledge about which actions are associated with which types of pain—though that doesn’t exactly break the utility-based models either.
Sure, but “rationality” and “values” are pretty orthogonal ideas. You can use rational thinking to pursue practically any set of values. I suppose if your values are crazy ones, a dose of rationality might have an effect.
Yes indeed. That’s been going on since the stone age, and it has left its mark on human nature.
Pretty much, but I think not totally. But we’ve gone far enough afield already. I’ll note this as a possible topic for a future discussion post.