HyperGraph Mutation Clarification (Potential Documentation Gap)

[MortimerE]

Hi everyone, we looking into the asmoses code and I had a few points of confusion I'd like help clarifying. I couldn't find explicit references to the mechanism behind changing the graph structure between exemplars in the meta population but from what I gather there's 2 algorithmic cases, one for boolean satisfiability problems and another for continuous function approximation.

For the Boolean case, it seems like deme opening is not mutation but a structured enumeration that samples feature pairs in canonical form. The graph structure evolves via flipped clause insertion with knobs controlling whether new literals actually materialize. Stochastic properties are introduced via uniform sampling of pairwise clauses via _perm_ratio, and disc_probe acts as a compile-time algebraic filter for redundancy pruning.

For the continuous case, the expansion of the graph exemplars is achieved by applying an algorithm that imposes a fixed structural expansion pattern based on preset flags. This seems thematically like a recursive template that seeks to create complex macrostructures by wrapping the existing exemplar in a pattern and then optimizing/pruning out subgraphs via the inner-loop neighborhood expansion + search. This seems somewhat deterministic and recursive as opposed to a more classical genetic algorithm that would sample a mutation/crossover etc. event using random or statistically distributed properties.
There does seem to be a form of crossover but as far as I can tell this only happens in the inner loop/deme expansion and not during metapopulation expansion, so it doesn't appear to be used to evolve the structure of the graphs themselves.

This leaves me with three general questions:

• Is my understanding of the mechanisms correct in terms of structural evolution of the graph exemplars in the metapopulation, which occurs sequentially between rounds of deme selection, opening, expansion?
• Is there a man or wiki page that I overlooked or is there a gap in documentation regarding the mechanisms at play for structural evolution and the best resource is the codebase itself?
• Are there specific reasons this approach was chosen over something more closely resembling mutation and crossover of the exemplars between generations?

[linas]

Hi Edward,

There are multiple wiki pages for moses and asmoses, as well as Moshe's original PhD thesis that explains the core concepts and structure. Start with https://wiki.opencog.org/w/Moses -- it has pointers to additional documentation.

Moses implements both conventional traditional mutation, as well as cross-over, but the mechanisms might not be particularly visible, as they are hidden underneath several layers of optimizations for speed, memory and representation. They're entirely "classical", there aren't any concepts here that wouldn't line up with conventional textbook material. What is new is that it is a "complete" system, optimized for performance and generality.

The only difference between the current code base and the PhD thesis is that the Bayesian optimization system was entirely ripped out; it was extremely slow; the three-way combo of mutation, cross-over and hillclimbing was about 100x or 1000x faster than attempting to maintain Bayesian estimates. Of course, that was then; it might be worth revisiting the Bayesian optimization, given that we now have technology like GPU's, and the well-articulated theory of recurrent NN's that could possibly overlap that of Bayesian optimization. Maybe. It would be a research project.

There's more than a few interesting research questions in here that have never been pursued; technologies like transformers have so thoroughly wiped out the competitions that more or less no one wants to go back and look at older techniques (such as this one: genetic programming) to explore different ways of doing things. In this sense, your interest is highly unusual.