Page 77 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

Days- Weeks Days- Months Mins months Secs- Months Mins Secs- Months Mins

Figure 5: From a manually controlled network to a cognitive autonomous network

able to add, remove or modify (e.g., split or combine) the
network configurations based on the learning of how or if the
network configurations are useful.

It is obvious that without the specialized DApps, the CE is
the default automation engine. The critical aspect of the
design is that the system does not have to observe all states
in order to know how to behave. Full knowledge of all states
can be achieved by extrapolation from known data, i.e., the
observations made in some states can be used to predict how
to behave in other network states.

4. LEARNING NETWORK BEHAVIOR – A CAN
EXAMPLE FOR MOBILITY MANAGEMENT

Let us assume a CAN and the use case of managing terminal
mobility. Using a RL DApp (e.g. like that in [10][12]), the
CAN is able to learn the optimal behavior for a mobility state
(the average velocity observed in the cell) as the best
Hysteresis (Hys) and Time-to-Trigger (TTT) values for the
state. For the network, the state-action space is very large and
Figure 4: CAN framework – Functions of CAN system not all states are frequently observable (some states are very
and related Cognitive Functions rare) yet the CAN needs to know how to behave in all states.
We show here that CANs can effectively learn the network
response in the unknown states based on their observations
The CME is the controller for the entire cognitive system
instance. First it decides if a DApp configuration in the known states. The CAN is particularized as supervised
recommendation should be enforced on the network. In so learning algorithms which, using multiple examples, learn to
doing it is responsible for concurrency control which is perceive a data element as instance of the earlier-on learned
different from logical coordination that is undertaken by the network’s response to mobility management parameters.
CE. The CME communicates the selected network Then, by reasoning over the DE’s relationship with the
configurations to the DAPPs and the CE (for their learning earlier examples, the algorithms can predict the likely
and coordination purposes) through interface f. Secondly, it network’s response to that instance.
also defines and refines the legal/acceptable candidate
network configurations for the different contexts of the 4.1. Simulation set up
different DApps based on the meta learning from the CE. In
the simplest form, the CME masks a subset of the possible For this study, we obtained data from simulations for 6
network configurations as being unnecessary (or velocity states [3,10,30,60,90,120 in kmph] and different
unreachable) within a specific abstract state. So, the set of mobility configurations (Hys and TTT) in an LTE network
possible network configurations is fixed and the CF/DApp of 21 states using a simulator described in [10][12]. For each
only selects from within this fixed set when in the specific configuration, we simulate 200 minutes of network operation
abstract state. A more cognitive CME may, however, also be and note the rates of handover (HO) successes, Ping-Pongs,

– 61 –