Page 75 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

Actions may be distinguished between automatic and con- areas of supervised, unsupervised, or semi-supervised
scious actions, respectively related to automatic and learning. Therein, the respective algorithm is given raw data
controlled processes [8]. “Automatic processes are inevita- from which it captures insight about the perceptions
bly engaged by the presentation of specific stimuli inputs, (concepts and contexts) of the data and the relationships
regardless of the agent’s intention”. Correspondingly, “... an thereof. Correspondingly, these ML techniques offer an
automatic process is modelled after the reflexes, taxes, and aggregated combination of perception and data organization.
instincts from physiology…” [8]. Note that the automation is Their perceptive capabilities are however limited, since they
not always innate but can also be achieved through extensive typically need a preprocessing (transduction) step that
practice. Meanwhile automatic actions can also “consume translates the observed natural data into a format over which
attentional resources once invoked by appropriate stimuli the algorithms are able to reason. Note also, that their ability
conditions” [8]. This explains the presence of subconscious to predict (e.g. through regression) may be interpreted as an
processes and actions as distinct from reflex processes and inference capability. This, however, is also limited by the
actions. Action automation also exists in organization and need for a post processing step required to interpret such
inference, albeit to a lesser degree especially in inference. predictions into the desired knowledge, decisions, or actions.
Here, the agent instead uses its cognitive functions to process
the available information and generate the optimal action. 2.4.3. Reinforcement Learning (RL)

2.4. Cognitive capabilities of AI/ML tools RL is the modern form of optimal control focusing on
learning how to optimally act in a given environment. It
AI and ML provide many techniques that process input data assumes abstractions of the states in the respective environ-
to generate solutions as knowledge, decisions, or even ment to which it learns the optimal actions. Evidently, RL
actions. For their use towards cognitive autonomy in net- offers mainly data organization and inference based on
works however, it is necessary to define the extent to which predefined abstract states.
these techniques achieve the processes in the perception-
reasoning pipeline model. For brevity, this section describes The capabilities of the three categories of AI/ML techniques
these capabilities without detailing the design or even show that to realize full cognitive autonomy, a system must
operation of the respective AI/ML techniques. The tools can combine multiple techniques each of which only achieves
be grouped into three major categories: (1) Classical AI subparts of the data processing pipeline. This understanding
systems which are concerned with figuring out a way to act is used in designing the framework for cognitive autonomy
in an environment; (2) Learning from data where the system in network management as described in the next section.
attempts to capture insights from a large data set; and (3)
Online learning where the system learns while acting in an 3. COGNITIVE AUTONOMY IN NETWORKS
environment that trains the system through feedback. The
recently famous fields of Neural Networks and Deep 3.1. Taxonomy - levels of automation in networks
Learning (or simply shallow and deep Neural Networks,
respectively) are computation models that allow for the Based on the forgoing discussion, the two dimensions for
efficient realization of learning solutions. describing an entity are its degree of independence in acting
and its level of intelligence in decision making. In the context
2.4.1. Classical artificial intelligence of CAN, we describe entities as automated, self-organizing,
cognitive, or autonomic as follows:
Classical AI techniques mainly refer to Expert, Closed-loop
control, Case-based Reasoning, Fuzzy Inference, and other  An automated entity is said to have a fixed way of
similar systems. Such systems are made up of two basic parts: behaving – responding to a certain stimulus - with the
a knowledge base and an inference mechanism. The behavior defined a priori by the entity’s creator
knowledge base holds semantic knowledge of objects  A Self-Organized (SO) entity is one capable of
including names of the objects and the known concepts, selecting actions as triggered by a given signal without
theories, and relationships thereof. The inference mechanism external control. This differs from an automated entity
includes procedures which examine the knowledge base in in that the internal mechanisms of such an SO entity are
an orderly manner and those used to reason and subsequently of no interest to its owner or user.
answer questions, solve problems or make decisions within  An autonomous or autonomic entity has the freedom to
the domain. It is evident therefore that they mostly achieve act as it wishes without influence from external entities
inference based on the relationships prescribed by the
knowledge base, although they may, to a limited extent,  A cognitive entity is one capable of making perceptions
afford some organization by learning new relations among and interpretations of its sensory inputs to derive
the data in the knowledge base. decisions through reasoning.

2.4.2. Learning from data In effect a self-organized entity may be cognitive but if
otherwise, such self-organization is achieved through a hard-
Learning from data includes the ML techniques in which a wired decision logic within the entity. Subsequently, all
system is trained from a trove of data – specifically the three cognitive entities are self-organized but not all self-

– 59 –