 | Backpropagation Class |
Inheritance HierarchyNamespace: RavingBots.MagicGestures.AI.Neural.Classic
Assembly: Assembly-CSharp (in Assembly-CSharp.dll) Version: 0.0.0.0
Syntaxpublic class Backpropagation
The Backpropagation type exposes the following members.
Constructors| Name | Description | |
|---|---|---|
 | Backpropagation |
Construct a new instance with the given RNG.
|
Properties| Name | Description | |
|---|---|---|
 | CurrentSettings |
The settings used in the current training.
|
| CurrentStatus |
The current status of the training.
|
| CurrentThread |
The handle to the training thread.
|
Methods| Name | Description | |
|---|---|---|
  | AdjustWeights(Layer, Single, Single, Single) |
Adjust the weights of every neuron.
|
| AdjustWeights(Perceptron, Single, Single, Single) |
Adjust the weights of the given neuron.
|
| CalcError |
Calculate new error values for the output layer.
|
| PropagateError |
Propagate error values from the output layer backwards.
|
| ResetNeurons |
Reset every neuron on every layer in the given set.
|
| RunAsyncTraining |
Start training the given network.
|
| RunTraining |
Run the backpropagation algorithm.
|
| StopAsyncTraining |
Request the training in progress to be cancelled.
|
| TrainSample |
Run a single iteration of the backpropagation.
|
Fields| Name | Description | |
|---|---|---|
 | _currentStatus |
The current status of the training.
|
| _mlp |
The network being trained.
|
| _samples |
The training set.
|
 | Random |
The random number generator used in the training.
|
RemarksWe train the neural network using the backpropagation algorithm. For every sample in the training set the input values are propagated forwards through all of the layers (Process(Single), then an error value is calculated (CalcError(MultilayerPerceptron, Single)) and propagated backwards from the output layer to the input layer (PropagateError(MultilayerPerceptron)), and finally weights on the neurons are adjusted (AdjustWeights(Layer, Single, Single, Single). This is the stochastic mode of learning, where every propagation results in the weight update.
See Also