Abstract:
The NNC (Neural Network Controller) automaticallv tunes the buffer size at the user/server level to eliminate anv chance of overflow in the
client/server interaction over a TCP logical channel. Together with the buffer tuning operations at the system/router level (e.g. the AQM (Active
Queue Management) activities) they form a unified solution. The power and stability of the NNC was verified over the Internet, but the result shows
that the drawback of the NNC is its long control cycle time. This drawback hinders the deployment of the NNC in the real-time applications. To
overcome this we propose the novel HRP (Hessian Based Pruning) optimization technique. This technique operates ,as a renewal process, and within
the service life of the Optimized NNC (O-NNC) the optimization operation repeats as renewal cycles. The feed-forward neural network configuration
of the O-NNC is optimized in everv cycle that involves two phases. In its original un-optimized form the NNC runs as a twin system of two
modules: "Chief + Learner". The O-NNC always starts with the unoptimized configuration. In the first phase the weights for the Learner's neural network
arcs are computed and sorted. Those arcs with weights insignificant to the control convergence speed and precision are marked. The marking
is based on "dynamic sensitivity analvsis" that utilizes the HBP technique. In the second phase the Chief optimizes the neural network by excluding/
skipping the marked arcs. The aim is to shorten the computation for the control cycle. The "HBP+NNC" is the basis of the O-NNC model, which
essentially uses virtual pruning because the marked arcs are excluded from the computation but not physically removed. While the Chief is carrying
out actual dynamic buffer tuning the learner undergoes training. The O-NNC model is verified by running the Java-based prototype on the Aglets
mobile agent platform in the Internet environment. The results are positive and indicate that the HBP technique indeed vields a shorter O-NNC control
cycle time than the original un-optirnized NNC in a consistent manner.
