lspector / taggp Goto Github PK

I thought I'd use this "issue" more as a discussion.

If you take a look at the newly refactored tags.clj and recursion.clj, you'll see that recursion.clj redefines all of eval-with-tagging. I had tried to abstract out all of the consequents of the conditional, so that eval-with-tagging for each condition would be calling

(cond (pred-fn? expression) (eval-pred-fn expression tag-space steps default-value)
      (another-pred-fn? expression) (trampoline eval-another-pred-fn expression tag-space steps default-value)
...

where the trampolined expressions replaced the recur call. Unfortunately this lead to stack over flows in the else clause, presumably due to the accumulation of calls on the stack from the un-trampolined call in its loop. I thought this might be a more elegant way to go, so instead of copying and pasting over the entire eval-with-tagging function, we could just modify the constituent eval functions (thereby also limiting the number of newly introduced bugs). However, to make this work, I would need to rethink the recursive calls in eval-with-tagging to handle the evaluation with args (presumably by carrying in it some extra variables to hold the deferred function and a queue of its args).

Anyone have any better thoughts on this matter?

Should recursion elimination be specified as a global? As an argument to evolve? The type is currently hardcoded in eval-with-tagging.

Create a new namespace to store generation data to be called in in core.clj. This will include things like the best individual, average error, etc. New features can then be added in an experimental namespace in the style of recursion.clj without having to alter the entire evolve function.

Default value version needs testing (in refactored branch).

Change taggp to properly handle untagging-based recursion elimination, both with real untagging and with default replacement.

Add system-wide tests in addition to unit test for new features.

Unit tests can be attached to functions in metadata (I'm currently having difficulty pulling this off).

Where to put larger system tests?

Implement random-seed hack thingie so random is truly random.

So far only a random-seed global has been added to the globals namespace.

Argument checking and parsing is currently done in the example files. This should be taken care of in core.clj

Update remaining examples to work with refactored taggp.

Rewrite taggp so that it no longer eats stack frames and dies for deep recursion.