Useless try to create neural network.
Instead of using gradient descent for learning my neural network I've tried simplier method. I tried to do the following:
layer[j]=layer[j]+activity[i]*sign(layer[j])*reward*lr
where layer[j] is weight of ith input of each neuron. activity[i] is latest value of that input. i is number of input. So it increase the absolute value of weight if action was rewarded and decrease it if not. But unfortanetely it doesn't work.
This network was tested on LunarLanderContinous-v2 from gym library. On average it do 1 landing over 1000 iterations.
python rl1.py
NOTICE: gym and Box2D libraries are required to run. Python version>=3.6