Category: Uncategorized

First issue with the rewriting system, the different fractals become more and more recurring until I knew almost all the fractal types my rewriting system had to offer (as often like a gaussian distribution). I proceeded to the analysis of the rewrited part of the fractals. Here an example of the few first line:

-ZbCCbZ*
-Zb/-//-/bZ*
-Zb/-//-/bZ*
-Zb/-//-/bZ*
-Zb/-//-/bZ*
-Zb/-//-/bZ*
-Zb/-//-/bZ*

This is a rewriting rule that has evoluted using my symmetry score algorithm. It is designed to give higher score to rewriter with higher symmetry. The system itself tend to give simple results, more often not using any of the complexity provided by the rewriter. Since it was such a waste, I tried forcing it to complexify. The technique I use were rule entropy, rule information compression, rule change rate by comparing each lines,… It seam that it is much harder to force the system to give up using the simplest rules. I liked some fractal that rules were symmetric, so I thought it might be a great idea.

fn = -rwevaluatorescore*entropy – (positiveloss – negativeloss) * entropy – entropy/2 (minimisation function)

Here the idea is to force the genetic algorithm to generate more complex fractal by giving an advantage to formulas with greater change in there structure (rwevaluatorescore). The process in mainly influenced by an entropy staircase (to gen algo always favor fractals with higher entropy). An importance is added to group of class of fractal (auto-encoder classifying NN net) and removed from another group of class. It’s hard to say if the NN has much influence there because the process is mostly entropy driven.

Now blocking all lower entropy results loss function only entropic and favoring complexity of the rewriting formula : MinLoss = -rwevaluatorescore/10 – entropy/2