Next test. I'm going to change all system to digital without use Opamps.
error propagation like cardiogram :D.
I can't up velocity because I need wait the signal be stabilized but it is slow because is obtained from low pass filter after PWM.
If period is reduced the stabilization is more fast but I loss precision and entire period is shared between 5 PWMs too.
Noise is natural regularization :)
corrections...
Limiters added...
modifications to limitate voltage before PIC input to avoid some undesired spikes.
PWM changed from 128Hz to 600Hz.
Preparing opamps and another PWM for neuron weight cost.
each weight has adjusted with ±100Mv max (output*weight).
Error propagation (yellow) of output neuron after propagation and cost calculation of memorized state0 (blue when yellow HIGH). Or no error propagation (LOW yellow) when state1 (actual input) is propagated (blue when yellow LOW)
On next image reward is 0 but with max/min rewards is reached from 0 to 5v.
ready cost calculation
Master neuron (first neuron on input layer) doing a propagate signal to adjacent neuron (first layer only) and all first layer neuron alternates between memorized state (state0) (HIGH) or respond to actual input (state1) (LOW) in a consecutively mode.