Ø Cognitive Packet
Network (CPN) (continued)
Ø
ØThe decisional weights of a RNN are increased or decreased
based on the observed success or failure of subsequent SPs to achieve the Goal.
ØGiven a goal G, reward R is: R = 1/G and RNN weights
updates are based on a threshold T:
Ø
Ø where
Rk, k = 1, 2, ... are successive measured values
of reward R and α is some constant (0 < α < 1) that is used to tune the responsiveness
of the algorithm: for instance α = 0.8 means that on the average five past values of R are being taken into
account.
ØNeurons are rewarded or punished based on the
difference between Rk, the current reward, and Tk-1, the last threshold.
ØIn CPN QoS metrics can be combined, e.g. the hop count
metric can be combined with the forward delay, so that the goal takes
into account both the length H and the delay D of the path.