Page 73 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

that are created based on fixed rules and input from Mobile eventually leads to the observed outcome, i.e., “the animals
Network Operator (MNO) and SON vendor. use their intelligence and experience to guide them” [6].

SON is however limited since the control loops are inflexible 2.2. Copying from human cognitive autonomy
to match the infinitely many contexts in networks. Thus, SFs
are unable to adapt to major environmental or operational According to various dictionaries, cognition relates to the
changes, from say technical upgrades, network architecture collection of mental sub-processes that support the process
modifications, or new operator business and service models. of acquiring knowledge and understanding through sensory
Such changes will occur even more often in 5G networks, so stimuli, experience, and thought. The human brain, which is
a more flexible and adaptive NM system is required. the best example of a cognitive entity, continuously executes
Cognitive capabilities derived from Artificial Intelligence the cognitive processes to process all the information
(AI) and Machine Learning (ML) promise better automation received from the environment and adequately analyze a
in networks, but the path thereto needs to be drawn out. This situation to flexibly adapt to its reality, demands, and
paper proposes such a framework and related models. changes. The mental processes and subsequent skills can be
grouped into [7]: basic, mutually independent processes that
The rest of the paper is organized as follows: Section 2 are independently fundamental to the functioning of the
reviews the notion of cognitive autonomy and describes a brain and higher or complex processes that are built up from
cognition model for the optimal design of Cognitive Auton- combinations of basic and other complex processes.
omous Networks (CAN) and functions thereof as well as the
extent to which AI and ML techniques achieve the desired The four basic processes are sensation, perception, attention,
functionality. Using the intuition from Section 2, Section 3 and memory. Sensation is the awareness of the various stim-
presents an AI/ML CAN implementation framework that is uli in our environment while perception enables us to process
applicable to cellular network environments. Section 4 then the received impulses and make out meaning from the stim-
presents the usage of such a framework towards learning the uli. Attention enables us to voluntarily and/or involuntarily
network’s response to different mobility states and configu- chose what we focus on amongst all the infinitely many con-
rations before Section 5 makes some concluding remarks. current stimuli from the environment. Finally, memory
allows us to encode the data that we either receive from the
2. COGNITIVE AUTONOMOUS SYSTEMS environment or mentally create, so that we can consolidate
and retrieve it at later points in time.
2.1. Automation, self-organization, and cognition
Higher processes presuppose the availability of knowledge
Automation has historically been equated with self-organi- which they put to use. Among these are thought, language,
zation, typically as derived from biologically inspired intelligence as well as combinations of these leading to
autonomy. A common example of self-organizing systems in problem solving and learning. Through thought, the brain
biology is the formation and management of colonies in creates concepts and processes them to derive new
social insects, e.g. haplometrosis in ant populations [3]. knowledge. Language enables us to produce and compre-
Many collective (social) activities performed by insects hend different sounds and words and to combine letters and
result in the formation of complex spatial-temporal patterns. phrases to express with precision whatever we want to com-
Without centralized control, workers are able to work municate. Intelligence, in all its different forms as
together and collectively tackle tasks far beyond the abilities intrapersonal, linguistic, logical-mathematical, musical, and
of any one individual. The resulting patterns produced by a the recently popular emotional intelligence, takes complex
colony are not explicitly coded at the individual level, but combinations of basic and other higher processes to manifest.
rather emerge from non-linear interactions between individ- Core to all these higher processes is the concept of reasoning
uals or between individuals and their environment [4]. A through which the brain combines different pieces of
good example here is the foraging for food by ant colonies information and knowledge to create new knowledge.
through which ants are able to establish efficient routes to
and from the nest to the food sources [5]. 2.3. Cognition: A perception-reasoning pipeline model

New studies, however, show that the behavior of colony From our analysis of above biological processes, we postu-
animals (especially ants) is more than random or determinis- late in the context of this work that cognition is a perception-
tic self-organization, i.e. “each foraging process of an animal reasoning engine which takes a piece of data, hereinafter
is also a learning process. With foraging repetition, long- called a Data Element (DE), and processes it to generate
term memory accumulates, an animal’s knowledge about the understanding and action(s). As shown in Figure 2, there are
environment of its nest gets richer, and the region that the four quasi-orthogonal processes of this engine that lead to
animal is familiar with continues to enlarge” [6]. The the realization of the two broad outcomes – perception and
observed emergent behavior is due to some internal cogni- reasoning. Perception prescribes the ability to make sense of
tive process in which the ants process information to make an incoming DE both on its own and in relation to data
decisions. Internally, within each insect, there are sub- elements about its context. On the other hand, reasoning
processes that the insect uses to compute its decision that implies understanding the DE and its implications which
then leads to the selection of the most appropriate actions for

– 57 –