Today I post on my youtube channel a new video about neural network and genetic algorithm test I did. Now , it’s about biped walk learning. I don’t touch the compiled code to achieve but more about which data I send to it. So all is about the python script in blender.
To achieve this result , I simulate what we call “recurrent neural network”. The type I did is a Helman type because it’s more simple and I don’t have to touch to the compiled code module. What I did is to take the calculated ouputs of the current tic ( frame ) and enter it has a inputs for the next tic ( frame ). So if before I have 12 inputs and 8 ouputs , now I get 20 inputs ( 12 + 8 ) and 8 ouputs. Here is a graphic of a Helman type :
Nice behaviors happen after 250 generations. After that, evolution is less visible. Here is the video of the result on biped :