Page 69 - Trust in ICT 2017

Trust in ICT 1

• Advanced control-based solutions

This technique requires complex models, with some unknown parameters (e.g., the machine
workload) that may be estimated online, to provide Adaptive Control. Adaptive Control requires an
identification mechanism and the ability to adjust controller parameters on the fly.
• Model-based machine learning solutions
This requires the definition of a framework in which to learn system behaviour and adjust tuning
points online. Neural networks are often useful to build a model of the world for control purposes.
Neural network solutions may be used to predict the system reaction to different inputs and, given
some training samples, to build a model. The structure of the network and the quality of the training
data are critical to performance. The accuracy of the results depend on these crucial choices, and
thus no a priori guarantees can be enforced.
Another model-based family of techniques is Genetic Algorithms. Using a genetic algorithm requires
selecting a suitable representation for encoding candidate solutions (in other words, a model). In
addition, some standard operators (crossover and mutation) must be defined and a mathematical
function must be provided to rate candidate solutions and select among them. The overhead of both
neural networks and genetic algorithms may in principle be very significant.

• Model-free machine learning solutions
This method do not require a model of the system. A notable example is Reinforcement Learning,
even if a recent research trend is to complement Reinforcement Learning solution with a model
definition. According to [58], Reinforcement Learning agents face three major challenges. The first
challenge is how to assign credits to actions, the second is how to balance exploration versus
exploitation and the third is generalization. The convergence time of a Reinforcement Learning
algorithm is often critical [26] and complementing them with a model of the solution space may
decrease it [59].
In summary, decision making is an essential functionality of ICT system. Apart from autonomic approaches,
trust based decision making solutions should be developed to provide more reliable and secure networking
and services.

8 Trust modeling and policy/rule-based decision making
There is a great diversity of trust models and they can be classified considering different features. However,
one of the aspects that takes more relevance, especially when one talks about testbeds, is the type of
information from which they compute trust. Some use experiences from previous interactions, some
opinions from other agents in the system, some analyse the underlying social network of agents or study the
information about the virtual organization to which agents belong, and even more complex examples exist.
Many combine several types of information to achieve better estimations.

8.1 Information context of a trust model
Information context denotes the sources of information and the flow of information from which a trust model
computes trust [60]. To graphically depict an information context of a general-purpose trust model, a schema
from [61] [62] can be build. The schema is shown on Figure 18 is centered on the agent that uses the trust
model, called agent a. It shows three information sources from which a`s trust model computes trust. The
agent can obtain information by interacting with agents, by asking for opinions, or by using information from
the environment.
Because the first two information sources are the most common in current trust models, it is highlighted
them and encapsulated other possible sources for trust computation in a special component called
environment; examples of such include the analysis of social networks, information about the virtual
organizations, etc.

61