This is a great introductory article, I heartily recommend it.
I read this myself few years ago, got very excited, and had a go at writing a program that could automatically perform the derivation from Nielsen's worked example:
This was really interesting to work on, but I guess I became disinterested in it after I got it to crudely work. Be aware that the quality of the code, documentation, and the entire idea of the "proof-search" aspect of it is probably not very good.
Pearl's "Probabilistic Reasoning in Intelligent Systems" was one of the most lauded books during my time in the university. I remember when Pearl gave a talk about his upcoming book Causality, my professor said that it might become one of the most important math and philosophy books of our life time. I never read Causality, and it is now 16 years since the publication. Can anyone who follows the field comment how it is perceived today?
Im not completely sold on his book being a textbook.
It is incredible bit of research, I agree, but it still has rough edges and should be treated as such, a work in progress. I do not think this book should be the bible of causality, and I think we're worse off for thinking it is.
Personally, I find the work of Richardson and Robins in Single World Intervention Graphs https://www.csss.washington.edu/Papers/wp128.pdf far more intuitive and compelling than Pearl's do-calculus. The notation is too slightly cleaner and it does away mostly with the unorthodox notation Pearl uses.
I would like to get an introduction to the main ideas in current causality research. Is Pearl's book still the best introduction, or would you recommend something else instead?
I think most people who study causal inference would say that Pearl's writing is seldom as clear as a person unfamiliar with his ideas would like it to be. These days, I suspect most people would benefit from starting with Morgan and Winship instead: https://www.amazon.com/Counterfactuals-Causal-Inference-Prin...
In addition to presenting Pearl's ideas more clearly than Pearl himself tends to do, Morgan and Winship also describe Rubin's work more fairly than Pearl does. (On the other hand, Rubin also tends to disparage Pearl's work more than is necessary.)
I'm 20 pages in. I'll eventually read it in full, but it is very math heavy and hard to digest. I am reading it at a very slow pace (a few pages per week) because that's the only way for me to assimilate the content.
My personal opinion is that it is a very good book, just a bit hard.
There will be a popular science book from him coming out, I read a draft a while ago. It will present it all in a digestible way.
The do-calculus to me looks a bit cumbersome. I'm not generally attracted to math that doesn't look like math, except if it is nice looking. :-)
On a more general level, when I'm building robots, I consider the interface between their minds and the world a Markov blanket. This means that state in the world can only be represented in the robot by learning through its interface by a combination of forward and inverse models. The existence of a physical barrier makes it possible to define information going through it. You might suggest that the world is created by the robot's mind, but that is probably difficult to back with any information criterion.
If the robot learns its body, it knows that it is itself who is performing a movement or if it is being pushed. Knowing that it was itself the actor is enough to know the direction of the causal arrow.
I read it a while ago and don't know that much about causality, but my conclusion was the opposite of the author's: that it's not possible to infer causal relationships from observations only. All the papers I have read that claim the magic phrase "causality just from observations" either redefine causality, or make some assumptions about what could be the cause.
The way it was done in the book required some tricks. And the way the tricks worked was to make assumptions about how the causality flowed, and that the model captured all possible causes of an effect. For example, with the "does smoking causes lung cancer" example (back propagation), you had to say that smoking causes tar in the lungs, which is the only possible source of lunch cancer. (this might not be exactly right, my memory is a bit hazy now). I'd be happy to hear from someone who is more of an expert than I am.
After reading that, I think that post actually supports my claim and makes me think the same thing after reading Judea Pearl's book. There is a redefinition of causality going on, and a bunch of implicit assumptions, namely that the structure of causality that is created is correct. Basically, if you can generate a bunch of nodes and arrows that are correct and the only possible causal explanations, then sure you can compute causality given these assumptions. The comments in that article point out these same problems.
If Correlation Doesn’t Imply Causation, then What Does? http://www.michaelnielsen.org/ddi/if-correlation-doesnt-impl...