README.md

In robertzk/monadicbase: The R Base Package

R with a history monad

The R base library equipped with a history monad will allow us to perform timeless computation. Specifically, instead of thinking of an R script as a "program" that "runs" forwards in time, we will be able to roll back to arbitrary points in time as if it was a single mathematical structure of which we were computing "time sections." These time sections are a view into the program (equipped with a set of inputs) and the union over time sections gives us full history of the program given a fixed set of inputs. This is equivalent to being able to retro-actively place a browser anywhere in the program, or run it backwards, and fully generalizes dump.frames, which merely offers a single time section.

The main advantage of this approach is that we implicitly are forced to track the dependency of every computation. However, as R operates on hierarchical heterogeneous structures (lists), we can perform a static analysis to deterine minimally parallelisable "substructure" computations.

In particular, if we are munging a dataset and feeding it to a statistical classifier, we should be capable of multi-threadedly parallelizing the computation of each variable in the final munged dataframe by performing a destructuring of the final program into a sequence of subprograms which together compose the original program, but are totally parallelized. Applied recursively to a large computation, this allows for the most general possible parallelization of subcomputations without losing the flexibility of R syntax and pliability.

The goal here is to develop an extension or re-formulation of R which is capable of performing just-in-time restructuring of arbitrary R programs into an isomorphic collection of R programs (collection, not sequence, since now by definition the order is independent) which are provably not further parallelizable.

robertzk/monadicbase documentation built on May 27, 2019, 10:35 a.m.