Alright another round of experiments for #wakegp has ended. This round was for determining if adding the conditional SelectP instruction to the set will have any effect on both program size and fitness.

Another round of experiments which is still going on, is determining how deletion mutation affects both program size and fitness, that is prevents bloat. Earlier experiments show that the simple parsimony pressure method I invented, has positive effect on both fitness and program size, but when there is also deletion mutation, the effect is even better. In those experiments, deleting an instruction randomly from all programs(that is a rate of 1.0) has positive impact on both fitness and average program size. Now I'm experimenting to see if more than one deletion per program would be even better.

#GeneticProgramming #GP #EvolutionaryML #ML #AI #MachineLearning #evolutionarycomputation #EC #LinearGeneticProgramming #wakeworddetection #wake_word_detection #optimization #opensourceML #opensource #FOSS #opensourceAI

Alright finished another round of #wakegp tests. It seems my invented method for parsimony pressure is effective and also has positive impact on fitness. This confirms findings or opinions of other researchers that bloat has negative impact on fitness.

Now started another round which will end in ~11h to see the effect of deletion mutation on fitness and program size, if any. Current experiments suggest that deletion mutation has no effect on either, not this side or the other side.

PS: I just hope when I want to publish the paper, not having a bachelor won't be problematic. However I don't have much of a problem about my paper getting rejected. Publishing the paper is a side effect with benefit. Results are my goals.

#EvolutionaryML #EC #AI #ML #MachineLearning #GeneticProgramming #LinearGeneticProgramming
#EvolutionaryMachineLearning #GP #bioinspiredcomputing #BioinspiredML #academia #academic #academics

I like productivity! So I've found the best #FFT size for #Wakegp. Some experiments about the #parsimonypressure method I have done shows effectiveness of my simple method. Now I'm experimenting on that to find out the sweet spot.

Regarding instructions, I already have done experiments. And the results were unexpected. Adding functions like Sine, Cosine, Logarithm and Square root have negative impact on fitness.

There are still a lot of parameters which I need to fine tune. Mutation rates, number of demes, tournment size, reproduction rate, crossover rate and a lot more. And I also need to experiment with #LexicaseSelection

After all these and when I have runs producing accurate enough programs, I should learn how to optimize the found programs to "summarize" them. I really hope compilers like #LLVM and #GCC already do so. I have doubts but I think most likely, they already have heuristics to summarize programs while keeping the logic.

God willing, when #repebble(the new #pebble #smartwatch) comes to market, I could try to see if I can use it there. Then #repebble would be the first industry application of my research.

cc @lspector

Edit: Regarding the simple parsimony pressure method I invented, I'm pretty sure already someone has invented it but I haven't found it in the literature. Just like tons of other things I invent or discover in Math or CS and then realize it has already been discovered or invented.

#MachineLearning #EvolutionaryMachineLearning #EvolutionaryML #ML #ArtificialIntelligence #AI #wakeworddetection #wake_word_detection #hotworddetection
#GeneticProgramming #LinearGeneticProgramming #MachineLearning

God willing, my #paper will be ready in about a month.

I'm gonna name it "WakeGP: Efficient and accurate wake word detection using linear genetic programming"

Also #GECCO 2024 is close. Not sure if it's better to present there or publish it a journal like #MIT #ECJ or #IEEE #EC

It is important for me that my paper will be accessible for everyone under a #free_content licence

#GP #GeneticProgramming #EvolutionaryComputation #research #academic #academia #LinearGeneticProgramming #ML #AI