Comment Deep learning should be easy (Score 2) 76
COBOL compiler option (ADATA) dumps all data structures and logic elements heuristics in a machine-readable format so the output of ADATA (to output stream SYSADATA, can be easily captured in a file.
What good is that? We loaded SYSADATA so âoeeverything the compiler knewâ was stored in RDBMS tables, graph DB would have been better perhaps. I used SAS but a fellow who worked for me wrote parser in COBOL.
It allowed unambiguous searches, and cross-references, ensuring corresponding data structures had same offsets and type & length & occurrences.
Point is, AI doesnâ(TM)t have to read source & parse it. Provably correct representation of logic data elements is readily available with perfect precision & accuracy, no ambiguity or inference needed. And in COBOL-land, when a new compiler version comes out, they âoere-compile the worldâ & run regression tests with pretty exhaustive aged & representative test data, ensuring all edge conditions fire. (Once burned, twice smart)
All this gives wonderful parallel regression tests bed, so reverse-engineering COBOL, even Object Oriented COBOL & forward-engineering into Some other language is pretty well testable. We did COBOL to Java to run on mainframe manually, & utility compared of same input has byte-equivalent & hash & crc verification of full equivalent output for batch jobs, after linkage editor & scheduler & SMF/RMF/DB2 traces all Included too.
But GIGO, youâ(TM)ve reproduced same bugs & need to maintain both manifestations for a bit as performance & thruput & concurrency & security & reliability & maintainability & etc. all are baked in.
So can & has been done. Should you?
Maybe move functionality to data as triggers; store procedures; constraints; user-defined types & functions or skip O-O step & go to declarative erlang for full parallelism?