In Solomonoff Induction the smallest program outputting observations up to the 
present point in time will continue to output _future_ observations as optimal 
predictions.

These future observations may be thought of as missing data for which optimal 
imputations are the output. 

In ordinary statistical imputation, there is a table of potential observations, 
not all of which have been made in all attributes (columns) of all cases 
(rows).  These are represented by a special symbol indicating no data (such as 
"n/a") in the corresponding cells.  In the case of a pure time series 
imputation, such as Solomonoff induction, the entire future is present as 
something equivalent to "missing data".*   

In the case of supplanting statistics used by the social sciences with 
Solomonoff induction, rather than a linear tape going off into the future, one 
has a spatial dimension representing simultaneity.  In this event model 
selection based on lossless compression must not disqualify a model 
(self-extracting archive of minimal length) for failing to reproduce the 
symbols representing missing data and, instead, producing imputations.


*An aside is that an intelligent observer is inducing a model of not just any 
future old future observations but a particular observer's decision process 
about what it will next observe, which means, interestingly, an imputed value 
function for sequential decision theory.  This means there is no way to escape 
from AIXI, even if one tries to strip off the sequential decision mechanism.

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T36c83eb0aa31fc55-Ma73be751a158f74717b7908b
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to