public Neighbour(
Random random,
int inputNodes,
int outputNodes,
int maxLayers,
int maxInnerNodes,
int[] layers)
{
this.random = random;
this.inputNodes = inputNodes;
this.outputNodes = outputNodes;
this.maxLayers = maxLayers;
this.maxInnerNodes = maxInnerNodes;
this.layers = layers;
Network = new MLPNeuralNetwork(layers);
}
public Neighbour(
Random random,
int inputNodes,
int outputNodes,
int maxLayers,
int maxInnerNodes)
{
this.random = random;
this.inputNodesEnabled = new int[inputNodes];
for (int i = 0; i < inputNodes; i++)
{
if (random.NextDouble() > 0.5)
{
this.inputNodesEnabled[i] = 1;
this.inputNodes++;
}
}
this.outputNodes = outputNodes;
this.maxLayers = maxLayers;
this.maxInnerNodes = maxInnerNodes;
var innerLayers = random.Next(maxLayers);
layers = new int[innerLayers + 2];
layers[0] = this.inputNodes;
for (int i = 1; i < innerLayers + 1; i++)
{
layers[i] = random.Next(maxInnerNodes);
}
layers[innerLayers + 1] = outputNodes;
Network = new MLPNeuralNetwork(layers);
}
static void Main(string[] args)
{
// Extract training data
var data = File.ReadAllText("C:\\Dev\\data.csv");
var jobData = data.Split('\n').Select(d => d.Trim());
jobData = jobData.Take(jobData.Count() - 1);
// Holds the extracted flowchart data.
var trainData = new List <float[]>();
// Holds the known classification / reinforcement values.
var reinforceValues = new List <float[]>();
// We only want to include distinct jobs to not skew our NN.
var existingJobs = new HashSet <Tuple <float, float, float, float, float> >();
// Iterate over the jobs and create the training & reinforcement data.
foreach (var job in jobData)
{
// We can safely parse all the values to floats.
var values = job.Split(',').Select(v => float.Parse(v)).ToArray();
// To mitigate skewing the NN from duplicate jobs, we need to use some indicator as a unique identifier.
// Population, ParcelCount, ItemCount, ModelSolidCount, StockpileSolidCount
var jobTuple = new Tuple <float, float, float, float, float>(values[0], values[1], values[2], values[3], values[6]);
if (!existingJobs.Contains(jobTuple))
{
existingJobs.Add(jobTuple);
// The last two values are our RAM & CPU usage, so dont include them.
trainData.Add(values.Take(values.Count() - 2).ToArray());
// Populate the reinforcenemt values by binning the RAM usage.
reinforceValues.Add(GetRamArray(values[14]));
}
}
// Define the layour for our NN.
// -- Input Nodes 13
// (population, parcelCount, itemCount, modelSolidCount, pitCount, stockpileCount, stockpileSolidCount,
// diggerCount, blend, npv, isSeeded, periodCount, seedPeriods, RAM, CPU_Usage)
// -- 1 Hidden Layer. 13 + 2/3 * 5 = 16 nodes
// -- 1 Output Layer. 4 Nodes (0-16, 16-32, 32-64, 64-128)
var inputNodes = 13;
var outputNodes = 4;
var innerLayers = 1;
var hiddenLayerDepth = 16;
// Inner Layer + Input + Output = 3
var layers = new int[innerLayers + 2];
layers[0] = inputNodes;
for (int i = 1; i < innerLayers + 1; i++)
{
layers[i] = hiddenLayerDepth;
}
layers[innerLayers + 1] = outputNodes;
// Create the NN.
var net = new MLPNeuralNetwork(layers);
// Train the network with our known good data.
net.Train(trainData, reinforceValues);
// Lets test out network with a known job.
var newJob = new float[] { 300, 7, 13, 4083, 3, 0, 0, 3, 1, 0, 1, 25, 6 };
var jobTest = net.FeedForward(newJob);
// Classify the result into the correct bin.
var classification = Classify(jobTest.ToList());
}
