Lua Neural Network

NeuralNetwork.lua

ACTIVATION_RESPONSE = 1




NeuralNetwork = {

  transfer = function( x) return 1 / (1 + math.exp(-x / ACTIVATION_RESPONSE)) end --This is the Transfer function (in this case a sigmoid)

}



function NeuralNetwork.create( _numInputs, _numOutputs, _numHiddenLayers, _neuronsPerLayer, _learningRate)

  _numInputs = _numInputs or 1

  _numOutputs = _numOutputs or 1

  _numHiddenLayers = _numHiddenLayers or math.ceil(_numInputs/2)

  _neuronsPerLayer = _neuronsPerLayer or math.ceil(_numInputs*.66666+_numOutputs)

  _learningRate = _learningRate or .5

  --order goes network[layer][neuron][wieght]

  local network = setmetatable({

    learningRate = _learningRate

  },{ __index = NeuralNetwork});

  network[1] = {}   --Input Layer

  for i = 1,_numInputs do

    network[1][i] = {}

  end

  for i = 2,_numHiddenLayers+2 do --plus 2 represents the output layer (also need to skip input layer)

    network[i] = {}

    local neuronsInLayer = _neuronsPerLayer

    if i == _numHiddenLayers+2 then

      neuronsInLayer = _numOutputs

    end

    for j = 1,neuronsInLayer do

      network[i][j] = {bias = math.random()*2-1}

      local numNeuronInputs = table.getn(network[i-1])

      for k = 1,numNeuronInputs do

        network[i][j][k] = math.random()*2-1  --return random number between -1 and 1

      end

    end

  end

  return network

end



function NeuralNetwork:forewardPropagate(...)

  if table.getn(arg) ~= table.getn(self[1]) and type(arg[1]) ~= "table" then

    error("Neural Network received "..table.getn(arg).." input[s] (expected "..table.getn(self[1]).." input[s])",2)

  elseif type(arg[1]) == "table" and table.getn(arg[1]) ~= table.getn(self[1]) then

    error("Neural Network received "..table.getn(arg[1]).." input[s] (expected "..table.getn(self[1]).." input[s])",2)

  end

  local outputs = {}

  for i = 1,table.getn(self) do

    for j = 1,table.getn(self[i]) do

      if i == 1 then

        if type(arg[1]) == "table" then

          self[i][j].result = arg[1][j]

        else

          self[i][j].result = arg[j]

        end

      else

        self[i][j].result = self[i][j].bias

        for k = 1,table.getn(self[i][j]) do

          self[i][j].result = self[i][j].result + (self[i][j][k]*self[i-1][k].result)

        end

        self[i][j].result = NeuralNetwork.transfer(self[i][j].result)

        if i == table.getn(self) then

          table.insert(outputs,self[i][j].result)

        end

      end

    end



  end

  return outputs

end



function NeuralNetwork:backwardPropagate(inputs,desiredOutputs)

  if table.getn(inputs) ~= table.getn(self[1]) then

    error("Neural Network received "..table.getn(inputs).." input[s] (expected "..table.getn(self[1]).." input[s])",2)

  elseif table.getn(desiredOutputs) ~= table.getn(self[table.getn(self)]) then

    error("Neural Network received "..table.getn(desiredOutputs).." desired output[s] (expected "..table.getn(self[table.getn(self)]).." desired output[s])",2)

  end

  self:forewardPropagate(inputs) --update the internal inputs and outputs

  for i = table.getn(self),2,-1 do --iterate backwards (nothing to calculate for input layer)

    local tempResults = {}

    for j = 1,table.getn(self[i]) do

      if i == table.getn(self) then --special calculations for output layer

        self[i][j].delta = (desiredOutputs[j] - self[i][j].result) * self[i][j].result * (1 - self[i][j].result)

      else

        local weightDelta = 0

        for k = 1,table.getn(self[i+1]) do

          weightDelta = weightDelta + self[i+1][k][j]*self[i+1][k].delta

        end

        self[i][j].delta = self[i][j].result * (1 - self[i][j].result) * weightDelta

      end

    end

  end

  for i = 2,table.getn(self) do

    for j = 1,table.getn(self[i]) do

      self[i][j].bias = self[i][j].delta * self.learningRate

      for k = 1,table.getn(self[i][j]) do

        self[i][j][k] = self[i][j][k] + self[i][j].delta * self.learningRate * self[i-1][k].result

      end

    end

  end

end



function NeuralNetwork:save()

  --[[

  File specs:

    |INFO| - should be FF BP NN

    |I| - number of inputs

    |O| - number of outputs

    |HL| - number of hidden layers

    |NHL| - number of neurons per hidden layer

    |LR| - learning rate

    |BW| - bias and weight values

  ]]--

  local data = "|INFO|FF BP NN|I|"..tostring(table.getn(self[1])).."|O|"..tostring(table.getn(self[table.getn(self)])).."|HL|"..tostring(table.getn(self)-2).."|NHL|"..tostring(table.getn(self[2])).."|LR|"..tostring(self.learningRate).."|BW|"

  for i = 2,table.getn(self) do -- nothing to save for input layer

    for j = 1,table.getn(self[i]) do

      local neuronData = tostring(self[i][j].bias).."{"

      for k = 1,table.getn(self[i][j]) do

        neuronData = neuronData..tostring(self[i][j][k])

        neuronData = neuronData..","

      end

      data = data..neuronData.."}"

    end

  end

  data = data.."|END|"

  return data

end

function NeuralNetwork.load( data)

  local dataPos = string.find(data,"|")+1

  local currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

  local dataPos = string.find(data,"|",dataPos)+1

  local _inputs, _outputs, _hiddenLayers, _neuronsPerLayer, _learningRate

  local biasWeights = {}

  local errorExit = false

  while currentChunk ~= "END" and not errorExit do

    if currentChuck == "INFO" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      if currentChunk ~= "FF BP NN" then

        errorExit = true

      end

    elseif currentChunk == "I" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      _inputs = tonumber(currentChunk)

    elseif currentChunk == "O" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      _outputs = tonumber(currentChunk)

    elseif currentChunk == "HL" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      _hiddenLayers = tonumber(currentChunk)

    elseif currentChunk == "NHL" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      _neuronsPerLayer = tonumber(currentChunk)

    elseif currentChunk == "LR" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      _learningRate = tonumber(currentChunk)

    elseif currentChunk == "BW" then

      currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

      dataPos = string.find(data,"|",dataPos)+1

      local subPos = 1

      local subChunk

      for i = 1,_hiddenLayers+1 do

        biasWeights[i] = {}

        local neuronsInLayer = _neuronsPerLayer

        if i == _hiddenLayers+1 then

          neuronsInLayer = _outputs

        end

        for j = 1,neuronsInLayer do

          biasWeights[i][j] = {}

          biasWeights[i][j].bias = tonumber(string.sub(currentChunk,subPos,string.find(currentChunk,"{",subPos)-1))

          subPos = string.find(currentChunk,"{",subPos)+1

          subChunk = string.sub( currentChunk, subPos, string.find(currentChunk,",",subPos)-1)

          local maxPos = string.find(currentChunk,"}",subPos)

          while subPos < maxPos do

            table.insert(biasWeights[i][j],tonumber(subChunk))

            subPos = string.find(currentChunk,",",subPos)+1

            if string.find(currentChunk,",",subPos) ~= nil then

              subChunk = string.sub( currentChunk, subPos, string.find(currentChunk,",",subPos)-1)

            end

          end

          subPos = maxPos+1

        end

      end

    end

    currentChunk = string.sub( data, dataPos, string.find(data,"|",dataPos)-1)

    dataPos = string.find(data,"|",dataPos)+1

  end

  if errorExit then

    error("Failed to load Neural Network:"..currentChunk,2)

  end

  local network = setmetatable({

    learningRate = _learningRate

  },{ __index = NeuralNetwork});

  network[1] = {}   --Input Layer

  for i = 1,_inputs do

    network[1][i] = {}

  end

  for i = 2,_hiddenLayers+2 do --plus 2 represents the output layer (also need to skip input layer)

    network[i] = {}

    local neuronsInLayer = _neuronsPerLayer

    if i == _hiddenLayers+2 then

      neuronsInLayer = _outputs

    end

    for j = 1,neuronsInLayer do

      network[i][j] = {bias = biasWeights[i-1][j].bias}

      local numNeuronInputs = table.getn(network[i-1])

      for k = 1,numNeuronInputs do

        network[i][j][k] = biasWeights[i-1][j][k]

      end

    end

  end

  return network

end

Out of curiosity, I noticed your training loop, specifically line 168, you switch which the variables using to index the nodes and the weights, was this intentional?

I think in your case, this works out since you take care of the output layer and restrict the creation of the hidden layers to all have the same node count, so you kind of loop orthogonal to what I had expected initially. Anyway, thanks for sharing! Learned a bit from it.

Hey I wrote this! 😄

For historical purposes here's the original archived forum link

And here's the repository I hosted it in a while back
https://github.com/AdamDorwart/LuaNeuralNetwork

Re: loop indexing on L168. It's been a long time since I've worked through this code but I believe the indexing is correct (but confusing). In this context i refers to the current layer, j refers to the current neuron within the current layer (i), and k is the referring to the current output neuron in the next layer i+1. It looks like k changes meaning throughout the code so that's probably why it's not clear. The loop ranges (i = table.getn(self),2,-1, j = 1,table.getn(self[i]), k = 1,table.getn(self[i+1])) are key here.

With all of that we can interpret:

weightDelta = weightDelta + self[i+1][k][j]*self[i+1][k].delta

as adding the product of the output neuron's (self[i+1][k]) input weight with its delta to current neuron's (self[i][j]) weight delta. A slightly different way of saying that would be that we're propagating an output neuron's (self[i+1][k]) weight (self[i+1][k][j]) between a particular input neuron (self[i][j]), and delta backwards to that same input neurons delta. The data structure can be thought of as self[layer][neuron][inputWeight] which you can verify on L58-62

I don't know if this is still looked at, but there's a game that can run Lua and with some modification I would like to use this for an Ai? I can't find where to input though.....

To elaborate, how would you train this? Online Lua runners just return nothing- I have no experience whatsoever, could I have some help?

Well first things off, There's an error already in the code that you must fix.
The error is at line 257.