|
||||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | |||||||||
SUMMARY: NESTED | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |
java.lang.Objectorg.joone.engine.RTRLLearnerFactory.Node
protected class RTRLLearnerFactory.Node
A node. A node sits inside a joone layer and at a given index. It is initially part of z but later on we will order z into Z and then K will be the index into Z at which this node is to be found. A node that is not an input node can be in U and all output nodes are in T. The sorted Z is indexed by K and has as its first few elements those nodes found in U - thus we can share the K index between these two sets. In order to speed up processing, a bias node is also created and stored inside z and eventually in Z, typically in the final position in either z or Z. The bias node simply fires 1s all the time and is used to optimise the bias weight. It's derivative to anything is zero. Each node has weights attached to it. These weights carry the potential of other nodes into this one and are stored in the weights list. We are specifically interested in those weights that are firing from nodes in U and will store them in a separate list called I. Since we store the weights attached to each node inside the node itself, this then becomes the weight matrix. The matrix is generally sparse since a node is only connected to a few other nodes, and we gain some performance this way, at the cost of a lot of indexing, both here and in the weights themselves. Partly for performance reasons we access members in this class directly rather than via encapsulated messages.
Field Summary | |
---|---|
protected double |
bias
The bias if this is a bias node |
protected java.util.List<RTRLLearnerFactory.Weight> |
I
The list of weights that carry a signal into this node from a node in U |
protected int |
index
The index into the layer at which this node is found |
protected java.util.List<RTRLLearnerFactory.InitialState> |
initialStates
The list of initial states that fire into this node, not used at present but maybe for a future version in which the state is also optimised. |
protected boolean |
inU
True if this node is in U, false if not |
protected int |
K
The node's number in Z and U if it is in U |
protected Layer |
layer
The layer at which this node is found |
protected java.util.List<RTRLLearnerFactory.Weight> |
weights
The list of weights that carry a signal into this node |
Constructor Summary | |
---|---|
RTRLLearnerFactory.Node(double bias)
Create a new bias node, typically with a bias value of 1 |
|
RTRLLearnerFactory.Node(Layer layer,
int index)
Create a new node from a joone layer |
Method Summary | |
---|---|
double |
getDerivative()
Retrieve the derivative of the node |
double |
getValue()
Retrieve the current value of this node |
Methods inherited from class java.lang.Object |
---|
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait |
Field Detail |
---|
protected double bias
protected Layer layer
protected int index
protected int K
protected boolean inU
protected java.util.List<RTRLLearnerFactory.Weight> weights
protected java.util.List<RTRLLearnerFactory.Weight> I
protected java.util.List<RTRLLearnerFactory.InitialState> initialStates
Constructor Detail |
---|
public RTRLLearnerFactory.Node(double bias)
public RTRLLearnerFactory.Node(Layer layer, int index)
Method Detail |
---|
public double getValue()
public double getDerivative()
|
||||||||||
PREV CLASS NEXT CLASS | FRAMES NO FRAMES | |||||||||
SUMMARY: NESTED | FIELD | CONSTR | METHOD | DETAIL: FIELD | CONSTR | METHOD |