Instrumentation & Measurement Magazine 25-6 - 27

Table 1 - Deme's Local and Continental model parameters
LOCAL MODEL
Layer Name
Conv_1D_1
Conv_1D_2
MaxPooling
Flatten
Hidden layer 1
Hidden layer 2
Hidden layer 3
Hidden layer 4
Hidden layer 5
Hidden layer 6
Hidden layer 7
Output Layer
CONTINENTAL MODEL
Specification
Filter = 32, Kernel = 4, Strides = 2
Filter = 32, Kernel = 4, Strides = 1
Chooses the highest number within
the window
Layer Name
Conv_1D_1
MaxPooling
Conv_1D_2
Changes data shape from 2D to 1D MaxPooling
Neuron = 512, Relu
Neuron = 64, Relu
Neuron = 32, Relu
Neuron = 32, Relu
Neuron = 16, Relu
Neuron = 5, Relu
Neuron = 2, Relu
Neuron = 1
Flatten
Hidden layer 1
Hidden layer 2
Hidden layer 3
Hidden layer 4
Hidden layer 5
Hidden layer 6
Output Layer
After registering each player, a background JavaScript script
was run in Swarmio's client software to query Swarmio's portal
to retrieve a list of all servers. Then, the script would cycle
through that list and, for each server, it would send 11 packets
to the server and receive a reply back for each from the server,
measuring the time it took from when a packet left the client
until its corresponding reply was received at the client. We
took the average of those 11 values as the average RTT. The
results can be seen in the Swarmio dataset presented in [10].
Compared to these directly measured RTT values, we noted
that Deme's average accuracy was 74%, falling short of the required
90%. We therefore invoked the CT step to refine the
model and improve its accuracy.
In the CT step, to gain an intuition into why Deme's accuracy
decreased in the real world, we looked at the data
Specification
Filter = 64 Kernel = 4, Strides = 2
Chooses the highest number within the
window
Filter = 32, Kernel = 4, Strides = 1
Chooses the highest number within window
Changes data shape from 2D to 1D
Neuron = 64, Relu
Neuron = 32, Relu
Neuron = 32, Relu
Neuron = 8, Relu
Neuron = 4, Relu
Neuron = 2, Relu
Neuron = 1
distribution of Swarmio and KING. Fig. 5 shows the differences
in the data distribution between the two datasets. In
Fig. 4. The MLOps concept.
September 2022
addition, the average RTT was 280 ms and 210 ms for KING
and Swarmio, respectively. It is therefore clear that there are
noticeable differences between the two datasets, and the aforementioned
concept drift is quite visible especially in Fig. 5b.
This is not surprising because KING was collected in 2002,
while networking technologies and infrastructure have improved
since then.
To refine the model, we can either train it from scratch, or use
transfer learning [11] which uses the " knowledge " acquired by
a pre-trained deep learning model to kick-start training a new
model. Because it uses previous knowledge as the starting
point, transfer learning is usually much faster than training
from scratch. It therefore helps with rapid generalization and
improving performance by
using the previously acquired
knowledge with
an already learned task.
Ideally, the pre-trained
model's layers are frozen
and attached to new unlearned
layers such that the
pre-trained layers' weights
remain unchanged during
training, while the weights
of the new model will be
updated. This allows the
pre-trained model layers
to extract the features and
then pass them to the new
model's layers to be optimized
and learned for the
new task or fine-tune the
IEEE Instrumentation & Measurement Magazine
27
Instrumentation & Measurement Magazine 25-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-6
