Page 76 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

organized entities are cognitive. Correspondingly for
networks, the following definitions hold with the relative
differences illustrated by Figure 3:
 A scripted or script-controlled network allows
scenario specific execution of automation scripts. By
and large, the network is still human controlled, but
allows the scripts to be run autonomously for those
specific and very routine activities.
 A Self-Organizing Network (SON) does not only
automate the selection and execution of actions but
also interprets events under different context to
determine the cause-effect relations of these events
under different contexts.
 An autonomous network is one able to act on its own
i.e. it does not take dictation or rules from anyone but
may however not be able to reason its environment or Figure 3: The abstract levels of autonomy in networks
even be smart enough to make the best static decisions
(e.g. incorrectly interpret its internal rules).
 A cognitive network is able to reason and formulate 3.3. The CAN framework
recommendations for subsequent behavior, even
where the recommendations must be approved by a To achieve the levels of automation described above, we
human operator before their execution. have proposed the CAN framework of Figure 5 as the system
through which NM applications may be designed and
Based on these definitions, a cognitive autonomous network integrated into a single cognitive autonomous system for net-
has both cognition and autonomy, i.e., it uses reasoning to work automation. The framework has five modules to be
formulate recommendations for action and subsequently integrated, specifically the Network Objective Manager
independently executes the derived actions. The degree to (NOM), Environment Modelling and Abstraction Engine
which the network is cognitive or not is represented by the 7 (EMA), Decision Applications (DApps), Coordination
capabilities or levels of autonomy illustrated by Figure 3. Engine (CE), and the Configuration Management Engine
(CME). The framework applies both to fully-centralized and
3.2. The vision: outlook to eventual CANs to quasi-centralized environments (e.g. respectively the
The characterization of the different levels of automation traditional 3GPP network management [9] and controllers
hint at the stages through which network management auto- expected in the new (edge) cloud RAN [10]). All the
mation is expected to transit. This path, illustrated by Figure modules are learning entities, albeit learning from different
5 starts with manual control through automated and data and towards different objectives.
cognitive automated control to eventually a cognitive auton-
omous network. The manually controlled network may The NOM learns how to translate the operator’s goals and
undertake some automation e.g. using scripts to automate intents into technical objectives to be fulfilled by the entire
routine actions for operator convenience. The automated net- network management system. The NOM aggregates the
work, exemplified by SON, has human-designed control concepts of intent-based, policy-based, and objective-based
loops that undertake simple control through monitoring and optimization to derive the technical objectives communi-
(re-)configuration of network parameters, but the functions cated over interface a. The EMA learns to abstract observed
are configured by the operator to optimize their performance. network events and contexts (from interface b) into
consistent network states that are communicated to all the
The cognitive automated network adds cognitive capabilities other modules over interface c.
over and above the automation functions while cognitive
autonomous networks take full control by not only learning The DApps, the core of the Cognitive Functions (CFs), are
the actions but also the possible applicable contexts in the the specialized applications responsible for specific network
network. The current network deployments are mainly at the configurations and optimizations. Example DApps will
second level since they have deployed several automation target minimization of interference, optimization of mobility,
functions in the network which are still largely being reduction of energy consumption, or coverage and capacity
controlled by the operators. There are, however, publications optimization. Each such DApp learns to select the most
that push the boundary towards level 3 by proposing learning optimal action for a specific state as detected by the EMA.
functions that take the responsibility of configuring the auto- Note that a CF may exclusively be only the DApp or may
mation functions away from the operator. This has mainly also include other functionality besides the DApp.
been referred to as Cognitive Network Management (CNM),
but the eventual step requires an end-to-end design of the The CE learns to detect and resolve any possibly conflicting
CAN in a way that embeds the learning capabilities of network configurations recommended by the DApps. Such
AI/ML techniques into each part of the cognitive functions. coordination must be natively multi-vendor since the DApps
will be expected to be supplied by different vendors.

– 60 –