TIL: probabalistic programming
Nov. 21st, 2019 08:33 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
timelets“It turns out, fortunately, that we can combine the expressiveness of first-order logic with the ability of Bayesian networks to capture probabilistic information concisely. This combination gives us the best of both worlds: probabilistic knowledge-based systems are able to handle a much wider range of real-world situations than either logical methods or Bayesian networks. For example, we can easily capture probabilistic knowledge about genetic inheritance:
for all persons c, f, and m,
if f is the father of c and m is the mother of c
and both f and m have blood type AB,
then c has blood type AB with probability 0.5.
The combination of first-order logic and probability actually gives us much more than just a way to express uncertain information about lots of objects. The reason is that when we add uncertainty to worlds containing objects, we get two new kinds of uncertainty: not just uncertainty about which facts are true or false but also uncertainty about what objects exist and uncertainty about which objects are which. These kinds of uncertainty are completely pervasive. The world does not come with a list of characters, like a Victorian play; instead, you gradually learn about the existence of objects from observation.
...
The problem is that we directly perceive not the identity of objects but (aspects of) their appearance; objects do not usually have little license plates that uniquely identify them. Identity is something our minds sometimes attach to objects for our own purposes.
The combination of probability theory with an expressive formal language is a fairly new subfield of AI, often called probabilistic programming.4 Several dozen probabilistic programming languages, or PPLs, have been developed, many of them deriving their expressive power from ordinary programming languages rather than first-order logic. All PPL systems have the capacity to represent and reason with complex, uncertain knowledge.”
Stuart Russell. “Human Compatible."
for all persons c, f, and m,
if f is the father of c and m is the mother of c
and both f and m have blood type AB,
then c has blood type AB with probability 0.5.
The combination of first-order logic and probability actually gives us much more than just a way to express uncertain information about lots of objects. The reason is that when we add uncertainty to worlds containing objects, we get two new kinds of uncertainty: not just uncertainty about which facts are true or false but also uncertainty about what objects exist and uncertainty about which objects are which. These kinds of uncertainty are completely pervasive. The world does not come with a list of characters, like a Victorian play; instead, you gradually learn about the existence of objects from observation.
...
The problem is that we directly perceive not the identity of objects but (aspects of) their appearance; objects do not usually have little license plates that uniquely identify them. Identity is something our minds sometimes attach to objects for our own purposes.
The combination of probability theory with an expressive formal language is a fairly new subfield of AI, often called probabilistic programming.4 Several dozen probabilistic programming languages, or PPLs, have been developed, many of them deriving their expressive power from ordinary programming languages rather than first-order logic. All PPL systems have the capacity to represent and reason with complex, uncertain knowledge.”
Stuart Russell. “Human Compatible."
no subject
Date: 2019-11-22 08:22 pm (UTC)