search

jellespijker / VisionSoilAnalyzer

Soil analyzer using an USB microscope - for Linux PC and Beaglebone Black
4 stars 1 forks source link

When learning the Neural Net optimizes towards all negative output neurons #85

Open jellespijker opened 8 years ago

jellespijker commented 8 years ago

The fitness routine has to be rewritten for the Genetic algorithms

latex_1d54fbc2610f7ea81bab9265f228f962

This results in a optimization function that hangs on a vector of all -1 or 1

jellespijker commented 8 years ago

Extra bonuses added to stimulate the population members that show a complete intersect with the goal vector and a bonus when the positive neuron intersects with the goal. The Bonus is a multiplication of the number of output neurons.

render

jellespijker commented 8 years ago

Implement issue #86 A revolution against the elite upperclass

jellespijker commented 8 years ago

The problem probably lies in the fact that the NN currently predicts either a -1 or a 1 in each neuron. This makes it virtually impossible to move incrementally towards an optimum.

Let the neurons fire a continuous value not just a step function.

jellespijker commented 8 years ago

Tricky f..in' function... It convergences, but then it get stuck on 30 % error. I will park this and continue to work on it when I got reliable learning data.

One possible solution is to have two NNś one for the roundness and one for angularity.

jellespijker commented 8 years ago

It could be that issue #92 is related to this bug

jellespijker commented 8 years ago

Not related to issue #92