tag:blogger.com,1999:blog-6923947282926324208.post6912535938250884482..comments2017-08-16T10:40:25.090+02:00Comments on Entropic and Fractal Intelligence: Solved atari gamesSergio Hernandezhttps://plus.google.com/107797149522609875320noreply@blogger.comBlogger6125tag:blogger.com,1999:blog-6923947282926324208.post-759364443610151602017-07-07T10:50:04.357+02:002017-07-07T10:50:04.357+02:00The problem about searching in the "Turing sp...The problem about searching in the "Turing space" is how do you "jump" from one point to another, how to make a small perturbation on a Turing machine so the resulting one makes "sense", is still similar to the original one (in its input -> output behaviour)... we need a replacement for X = X+epsilon but on the code space, and we haven't any (please please please correct me if I am wrong!).<br /><br />Basically this space is too sparse to move around, the sensible algorithms are way to distant one from another, and the gap is filled with billions of bullshit variations. We need a way to walk this landscape without "stepping on a bullshit".Sergio Hernandezhttps://www.blogger.com/profile/18108694861191833007noreply@blogger.comtag:blogger.com,1999:blog-6923947282926324208.post-75017200345981217232017-07-04T18:38:20.955+02:002017-07-04T18:38:20.955+02:00Well, good methods to approximate incomputable SI ...Well, good methods to approximate incomputable SI would be an interesting research topic. A naive one would be Levin search, then there's Hutter Search and Juergen Schmidhuber's OOPS and Goedel Machine. However, these models are mainly of theoretical interest. In practice they mostly just solve some toy problems and fails to scale.<br /><br />Personally I think directly searching in the space of all Turing Machine tend is intractable. However the principle of searching for simple models could apply to other world modellers.Adam Beckerhttps://www.blogger.com/profile/18082529597907247625noreply@blogger.comtag:blogger.com,1999:blog-6923947282926324208.post-19094493353381147142017-07-04T16:53:44.907+02:002017-07-04T16:53:44.907+02:00"All possible programs" sounds too infin..."All possible programs" sounds too infinite to me!<br /><br />The idea of finding less complex program for a task is quite appealing, but if you should look for the "less complex program to find the less complex program in a general form"... what would you find? Would you need it before going into AIXI?Sergio Hernandezhttps://www.blogger.com/profile/18108694861191833007noreply@blogger.comtag:blogger.com,1999:blog-6923947282926324208.post-36914951847624196082017-07-04T16:18:35.441+02:002017-07-04T16:18:35.441+02:00Hmm, I haven't read all your blog post so far,...Hmm, I haven't read all your blog post so far, and I'm not an expert on AIXI either. Still, I'd like to point out a major difference. AIXI, or more generally, Algorithmic Information Theory (AIT) based methods, does provides a solution to world modelling problem, including partial observability.<br /><br />For example, you want to predict next outcome of binary string "1101001000100001 ...". In fact you totally have no idea about the underlying model. However this problem would be too easy for a human IQ test. You would say, it's "000001...". But why is that? Why not "010010"?<br /><br />The AIT solution: write down all possible C programs, including non-halting ones, that prints first 15 character as the observed sequence.<br /><br />Now, if we continue to execute all these programs, some of them will print a "0" while others will print a "1". How do we decide the probability of the next digit, given there are infinite many programs? Well, we take weighted average of these programs, where the weight of a particular program is 2^(-size_of_the_program_in_bits).<br /><br />This process is called Solomonoff Inductive Inference. In a short word, if a short program can represent the observation, it gets higher weight. We prefer simple models over complex ones.<br /><br />Since the program size, or Kolmogorov complexity, shares many property with Shannon's information entropy. I get the feeling Entropic Intelligence and AIT does complementary things: one maximize the entropy of future, while the other minimize the "entropy" of past.Adam Beckerhttps://www.blogger.com/profile/18082529597907247625noreply@blogger.comtag:blogger.com,1999:blog-6923947282926324208.post-71041030400377801652017-07-04T14:05:31.525+02:002017-07-04T14:05:31.525+02:00Hi Adam, I don't actually knew about AIXI, but...Hi Adam, I don't actually knew about AIXI, but after a quick search and read, it seems to be scanning the future outcomes making them to "fade away" in importance as they get more into the future (correct me if I am wrong!) so the nearer in time an event is, the more "decisory" it is.<br /><br />If that is the case, they both are similar but this "fading with time" heuristic is a "quick and dirty" trick -one that I already tried in my algorithm before going fractal- compared as how I do it now.Sergio Hernandezhttps://www.blogger.com/profile/18108694861191833007noreply@blogger.comtag:blogger.com,1999:blog-6923947282926324208.post-28549176351160821562017-07-04T11:39:48.011+02:002017-07-04T11:39:48.011+02:00I would like to know how well this performs agains...I would like to know how well this performs against AIXI (or more precisely, an approximation to incomputable AIXI). The CTW approximated AIXI is also a general algorithm without need to train internal parameters. It can also solve toy problems such as PacMan and has a very grounded, albeit different, theoretical basis.Adam Beckerhttps://www.blogger.com/profile/18082529597907247625noreply@blogger.com