Many people work in artificial neural nets late 1990. Are you one of them? The field where ann model will use called "machine learning" and now this field called deep learning where we study deep learning methods on various deep learning model. Initially a group of scientists build mathematical model ann.
Dietrich Dörner ann model little bit different then above. According to Dörner neurons are consist of
Activity "A"
Max activation Function "Max"
Amplification factor "Amp"
Threshold Value "t"
So many thing still unclear in start. After research I will understand all elements.
Activity "A" start from 0 to 1. Max activation function "Max" range 0 to Max. In some cases Max will be 1. Amplification factor "Amp" = Vout/ Vin and threshold from 0 to 1. 0.5 consider optimal threshold value.
Output Formula:
O = min(Max, Amp * A) ------------------------------------------------------------------------------ (1)
According to him neurons are four types,
Activator
Inhibitor
Associate
Disassociate
Activator and inhibitor neurons take part in creating, remove, change and activate their others neurons links. Inhibitor neurons are opposite. Activator are called excitatory.
Let assume a neuron "i" is an activator neuron and it will covey input after processing to neuron j for further processing then v[i,j] = O[i] * w[i,j]. Here O[i] is an output of neuron i and w(ij) is the weight of both Oi and successor neuron j. In opposite case v[i,j] = - w[i,j] * O[i]. If you want to calculate neuron Aj "Activation" then
A(j) = max(0, Σ v[i,j] - t {i = (0 to n)}) -------------------------------------------------------- (2)
Associate and Disassociate
Association and dissociation depend upon four further variables.
Learning Constant L
Dissociation constant D
Decay constant K
Decay threshold T
Associate Neuron transmit activation value to other neuron j and j itself also activated. Then weight of neuron j and all other activated neurons will be increase and new value of wij will be calculated
w[i,j]{new} = Sqrt(w[i,j]{old} + A[i]*A[j]*A[associate]* w[associate,j] +L))^2 ----------------- (3)
If neuron is not active then
w[i,j]{new} = Sqrt(max(0,w^2[i,j] - K)) --------------------------------------------------------------- (4)
if w[ij] < T then link are weeks, K value is quite small 0.05
Like inhibitory neurons disassociate are opposite of associate neurons.Disassociate neurons weight calculate through given formula
w[i,j]{new} = Sqrt(max(0, w^2[i,j]) - A[i]*A[j]*A[associate]* w[associate,j] -D) ------- (5)
Now you can create a simple feed forward for your agent. If you had an experience on ann then this information provide knowledge about black box.
Have a good day thank you