changed model of neural network

diff --git a/nn/data.lua b/nn/data.lua
index b05a2af669a188ef91d5880f8126ba780d19d35f..061bdaafe1df5f34ba9f1439bdb64f470ba83412 100644 (file)
--- a/nn/data.lua
+++ b/nn/data.lua
@@ -92,7 +92,12 @@ function saveNeuralNetwork(network, file_path)
 
        for l = 1, numLayers do
                local layer = network.network[l]
-               file:write(tostring(layer.bias) .. "\n")
+
+               for n = 1, #layer.neurons do
+                       file:write(tostring(bias) .. ' ')
+               end
+
+               file:write("\n")
 
                for n = 1, #layer.neurons do
                        local neuron = layer.neurons[n]
@@ -127,12 +132,15 @@ function loadNeuralNetwork(file_path)
 
        for l = 1, numLayers do
                log.log(log.LOG_INFO, "Reading layer " .. l)
-               local bias = file:read("*number")
-               network.network[l].bias = bias
 
                local upper = layerCount[l - 1]
                if upper == nil then upper = layerCount[l] end
 
+               for n = 1, layerCount[l] do
+                       local bias = file:read("*number")
+                       network.network[l].neurons[n].bias = bias
+               end
+
                for n = 1, layerCount[l] do
                        for w = 1, upper do
                                local weight = file:read("*number")

diff --git a/nn/nn.lua b/nn/nn.lua
index 102b98cdb064eb818f0065bcfb9ae3dd41f65931..c98608eb1bad75eb89e5534343e40623e812a1f4 100644 (file)
--- a/nn/nn.lua
+++ b/nn/nn.lua
@@ -5,18 +5,21 @@ function Neuron:init(input_count)
        self.value = 0
        self.delta = 0
        self.weights = {}
+       self.bias = math.random() * 1 - .5
        for i in range(1, input_count) do
                self.weights[i] = math.random() * 1 - .5 -- Initialize to random weights
        end
 end
 
-function Neuron:activate(inputs, bias)
+function Neuron:activate(inputs)
        local weights = self.weights
+
+       local activation = self.bias
        for i = 1, #weights do
-               bias = bias + (weights[i] * inputs[i])
+               activation = activation + (weights[i] * inputs[i])
        end
 
-       self.value = bias / (2 + 2 * math.abs(bias)) + 0.5
+       self.value = activation / (2 + 2 * math.abs(activation)) + 0.5
 end
 
 Layer = class()
@@ -28,8 +31,6 @@ function Layer:init(neuron_count, input_count)
        for i = 1, neuron_count do
                self.neurons[i] = Neuron(input_count)
        end
-
-       self.bias = math.random()
 end
 
 NeuralNetwork = class()
@@ -59,9 +60,8 @@ function NeuralNetwork:activate(inputs)
                        inputs[j] = prevCells[j].value
                end
 
-               local bias = self.network[i].bias
                for j = 1, #cells do
-                       cells[j]:activate(inputs, bias)
+                       cells[j]:activate(inputs)
                end
        end
 end
@@ -96,10 +96,11 @@ function NeuralNetwork:back_propagate(inputs, outputs)
        end
 
        for i = 2, numLayers do
-               self.network[i].bias = self.network[i].neurons[#self.network[i].neurons].delta * learningRate
                for j = 1, #self.network[i].neurons do
+                       self.network[i].neurons[j].bias = self.network[i].neurons[j].bias + self.network[i].neurons[j].delta * learningRate
                        for k = 1, #self.network[i].neurons[j].weights do
                                local weights = self.network[i].neurons[j].weights
+
                                weights[k] = weights[k] + self.network[i].neurons[j].delta * learningRate * self.network[i - 1].neurons[k].value
                        end
                end