Hi. I am giving a talk on machine intelligence next week. I have a slide about "generality" and I have an associated question that I thought I would try running by you guys.
My question is basically: what do you think of this presentation, and how can I improve on it? The presentation goes something like this: My slides refer to definitions of intelligence, that say that it is "general" and then mention that the "no free lunch" theorems of Wolpert and Macready from 1997 say that no algorithm performs better than any other on arbitrary search problems - and so an intelligent agent can't really be "general" in this way. I then argue that some kinds of assumption about the space of problems likely to be faced are actually quite reasonable and acceptable. I then list some of these assumptions: One is "the uniformity of nature". Physics appears to be fairly uniform in space and time. The uniformity of nature allows induction to work. It means that the past is relevant to the future. It means that experiences in one place are relevant in other places and that experiences at one time are also relevant later. Matter in the universe also exhibits a fair amount of regularity, due to repulsive forces, and the common, diluting forces of radiation, diffusion, turbulence and entropy increase. These make concentrations of matter tend to spread out. That is especially true far from deep gravity wells (where attractive forces dominate). That happens to be where most living systems find themselves. These effects compound the "uniformity of nature" effect. Another is Occam's razor. Occam's razor says that simpler hypotheses that are compatible with observations should be preferred. Occam's razor has subsequently been generalized to include a measure of time complexity. Another involves nature's preference for simplicity. One cause of this is locality. Physics is local (quantum mechanics notwithstanding). Locality imposes limits on the complexity of observed sequences, and it means that sequences with simpler generators are actually more likely to be observed. In three dimensional space, radiation and diffusion make distant effects less relevant. Again, these factors lead to simpler sequences being more likely to be observed. I then say other types of assumption associated with the laws of physics, regularity of the universe's contents, or the special place of living systems in the universe could also be possible. Lastly, I argue that these assumptions could be mistaken. Probably the most important example where this could happen is if your environment is being controlled by an adversary. They could then manipulate things so that your assumptions are systematically mistaken. -- __________ |im |yler http://timtyler.org/ ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T3326778943da25b8-Ma2d5ae6848b32eaada5d6497 Delivery options: https://agi.topicbox.com/groups/agi/subscription
