Download ЛЖЖЛ - Cognitive Systems

Transcript
194
CHAPTER 9. NEURAL NETWORK MODELS AND FUNCTIONS
9.13.3 ARTMAP
9.13.3.1 Structure of an ARTMAP Network
Since an ARTMAP network is based on two networks of the ART1 model, it is useful to
know how ART1 is realized in SNNS. Having taken two of the ART1 (ARTa and ARTb)
networks as they were dened in section 9.13.1, we add several units that represent the
MAP eld. The connections between ARTa and the MAP eld, ARTb and the MAP eld,
as well as those within the MAP eld are shown in gure 9.12. The gure lacks the full
connection from the Fa2 layer to the Fab layer and those from each Fb2 unit to its respective
Fab unit and vice versa.
Figure 9.12: The MAP eld with its control units.
The map eld units represent the categories, onto which the ARTa classes are mapped9.
The G unit is the MAP eld gain unit. The units rm (reset map), rb (reset Fb2 ),
rg (reset general), (vigilance) and d1 (delay 1) represent the inter-ART-reset control.
and qu (quotient) have to realize the Match-Tracking-Mechanism and cl (classied)
and nc (not classiable) again show whether a pattern has been classied or was not
classiable.
9.13.3.2 Using ARTMAP Networks in SNNS
Since the trainable weights of an ARTMAP network are primarily the ones of the two ART1 networks ARTa and ARTb , it is easy
to explain the ARTMAP initialization function ARTMAP Weights. To use this function
you have to select ARTMAP Weights from the menu of the initialization functions. For
ARTMAP Weights you have to set the four parameters a , a , b and b . You can look up
the meaning of each pair ? , ? in section 9.13.1.2, for the respective ART?-part of the
network.
ARTMAP Initialization Function
9
Dierent ARTa classes may be mapped onto the same category.