Page 66 - Trust in ICT 2017

1 Trust in ICT

incorporating system and environmental monitoring information into the genetic algorithm such that specific
changes in the environment automatically drive the evolutionary process towards new viable solutions. They
have applied this genetic-algorithm based approach to the dynamic reconfiguration of a collection of remote
data mirrors, with the goal of minimizing costs while maximizing data reliability and network performance,
even in the presence of link failures.
Furthermore machine learning techniques are often employed as decision mechanisms for a variety of
systems as it allows computers to evolve behaviours, based on empirical data, for example from sensor data.
Regarding this, a decision making system based on a neural network and a reinforcement learning algorithm
is discussed in [52].

Figure 15 – Neural network topology [52]

Martina et el implemented an artificial neural network, with the purpose of learning the best policy for
control. This means the neural network has to produce the next step control outputs from the current
situation, with the purpose of reducing the error between the measured heart rate and the desired one.
Every time we have a new sample, we feed that into the network and update its weights according to the
gradient of the error we are experiencing.

The network topology as shown in Figure 15, is composed by four different input sources, corresponding to
the desired heart rate, the actual heart rate and the two control inputs: number of cores and frequency. With
three neurons in the (single) hidden layer and two output neurons we learn the relationship between the
inputs and the (possibly optimal) control strategy. It is worth stressing that we didn’t train the network before
launching the experiments and the network itself is trained online, updating the weights according to the
experienced error with a gradient descent method.

Another alternative technique that can be applied to trust decision making process is use of reinforce learning
mechanisms as stated in [53]. Reinforcement learning is about learning from interaction how to behave in
order to achieve a goal. In here, learner is not told which actions to take, as in most forms of machine learning,
but instead must discover which actions yield the most reward by trying them. In the most interesting and
challenging cases, actions may affect not only the immediate reward but also the next situation and, through
that, all subsequent rewards. These two characteristics (i.e. trial-and-error search and delayed reward) are
the two most important distinguishing features of reinforcement learning.

58