Code proposals in regard to the course "Artificial Intelligence for Knowledge Discovery"

1. Three-Level XOR

mlp = nn.Sequential() -- container for the network

inputs = 2; -- input layer
outputs = 1; -- output layer
HUs = 20; -- hidden layer

-- create the network with hidden layer
mlp:add(nn.Linear(inputs, HUs))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(HUs, outputs))

criterion = nn.MSECriterion() -- loss function

Within threelevelxor.lua a small implementation of a neural network is provided for solving the 3-Level-XOR problem.

2. Monte Carlo Domineering

The file AIdomino.lua contains the implementation of the Domineering problem in Torch. It includes:

• the implementation of the neural network itself

local convnet = nn.Sequential()
  local nplanes = 64 -- hidden nodes
  local num_rows = 4000 -- total number of training cases

  convnet:add(nn.SpatialConvolution(3, nplanes, 5, 5,1,1,2,2))
  convnet:add(nn.ReLU())
  convnet:add(nn.SpatialConvolution(nplanes, nplanes, 3, 3,1,1,1,1))
  convnet:add(nn.ReLU())
  convnet:add(nn.SpatialConvolution(nplanes, nplanes, 3, 3,1,1,1,1))
  convnet:add(nn.ReLU())
  convnet:add(nn.SpatialConvolution(nplanes, nplanes, 3, 3,1,1,1,1))
  convnet:add(nn.ReLU())
  convnet:add(nn.SpatialConvolution(nplanes, nplanes, 3, 3,1,1,1,1))
  convnet:add(nn.ReLU())
  convnet:add(nn.SpatialConvolution(nplanes, 1, 3, 3,1,1,1,1))
  convnet:add(nn.View(64))
  convnet:add(nn.SoftMax())
  convnet:add(nn.View(8,8))

  criterion = nn.MSECriterion()
  error = 1

• the generation of training and test data (see dom.java and domineering.csv)
• the import of the training and test data (see csvreading.lua)

function M.loadData(dataFile)
  local dataset = {}
  i = 1
  local x = torch.Tensor(1000,3,8,8) -- input values

  local y = torch.Tensor(1000,8,8) -- output values
  for line in io.lines(dataFile) do
    local values = line:splitAtCommas()
    --print(values)
    count = 1
  	for v=1,3 do
  		for u=1,8 do
  			for t=1,8 do
  				x[i][v][u][t] = values[count]
  				count = count +1
      	end
  		end
  	end

  	for u=1,8 do
  		for t=1,8 do
  			y[i][u][t] = values[count]
  			count = count +1
  	  end
  	end

    i = i + 1
    if i==1001 then
    	break
    end
  end

• and the data augmentation (see dataaugmentation.lua)
  • for increasing the accuray of the neural network

via horizontal, vertical or a combined flip

function flip(matrix, type)
    local result = torch.Tensor(8,8)
    if type == 1 then
      result = image.hflip(matrix)
    elseif type == 2 then
      result = image.vflip(matrix)
    elseif type == 3 then
      result = image.hflip(image.vflip(matrix))
    end
    return result;
end

3. Results

This implementations achieves a good accuracy with a total error of approx. 0.015381567763701 after 4000 iterations.