The concept of convergence in numerical methods could be useful here. In the AI case, it would have a large number of models ranging from high-energy physics to, say, social sciences, and it would run a number of models in the neighborhood of the one looking most suitable for a particular problem. It will check that the solution is not very sensitive to increase in resolution, i.e. to applying progressively more detailed models.
If it finds a situation where there is a decision gap between two neighboring models, it will make an effort to fill in the gaps in its understanding of the world before returning to solving a specific problem and validating convergence given its new worldview.
In general, most of the sub-problems you find in any new research are not new and have been solved elsewhere, the hard part is to formulate them abstractly enough to be able to google the relevant concepts.
In general, most of the sub-problems you find in any new research are not new and have been solved elsewhere, the hard part is to formulate them abstractly enough to be able to google the relevant concepts.
The concept of convergence in numerical methods could be useful here. In the AI case, it would have a large number of models ranging from high-energy physics to, say, social sciences, and it would run a number of models in the neighborhood of the one looking most suitable for a particular problem. It will check that the solution is not very sensitive to increase in resolution, i.e. to applying progressively more detailed models.
If it finds a situation where there is a decision gap between two neighboring models, it will make an effort to fill in the gaps in its understanding of the world before returning to solving a specific problem and validating convergence given its new worldview.
That sounds like a concept I should look more into. Do you have any recommended references?
http://en.wikipedia.org/wiki/Rate_of_convergence
http://en.wikipedia.org/wiki/Numerical_methods_for_ordinary_differential_equations#Convergence
also http://en.wikipedia.org/wiki/Numerical_stability
In general, most of the sub-problems you find in any new research are not new and have been solved elsewhere, the hard part is to formulate them abstractly enough to be able to google the relevant concepts.
Thanks!
Very true.