shapedsundew9/erasmus-gp

Meta-genetic programming

erasmus-gp's Issues

Shadow Frameworks

Concept
A shadow framework is a package or library that can produce the same semantic result for a codon as its default definition. For example, tensorflow and numpy as in this ChatGPT example: https://chatgpt.com/share/7bbaced8-789d-4f0c-8323-e1d0f09f573f

The intent of this feature is to provide the optimization opportunities for the runtime.

Another example is vanilla python and sympy: A GC could (and often can) be radically reduced in runtime by simplifying the expression it represents: https://chatgpt.com/share/ef912723-cdc7-400c-83a1-3fee7a674152 or Common Sub-expression Elimination (CSE): https://chatgpt.com/share/7e3cfd91-170b-41fd-9557-2e482a90142e

Shadow Frameworks are only a runtime optimisation mechanism and the representation of the GC graph is not altered.

Impementation Thoughts
Shadow Framework support for the entire GC (including all sub-GC's and codons) could be represented in the properties field. i.e. they all have the property set. This lets the execution environment writer know the whole tree can be represented by the shadow framework.

The validated Shadow Framework implementation for the whole GC tree could be stored in the metadata of the GC.

There could be a 'custom' shadow framework i.e. a hand crafted validated & optimised version of the GC.

Validation must occur. It is recognised that precise semantic mapping is not a realistic goal. Order of operations, rounding, status side effects etc. will cause very tiny and most often irrelevant differences. Validation rules must be considered i.e. how exact is 'the same' and can be done automatically.

The validation state of a Shadow Framework for any GC is dynamic and validity must be versioned i.e. to optimized implementation in the metadata must be versioned. In this way bugs can be fixed.