Design cost and performance optimisation algorithm

data model to predict probabilities and optimised KPI alarms values. Algorithmic rules have been developed using our input variables (show in tables 5.1& 5.2) to estimate each of the four types of optimisation cost. Each of these costs are associated with the KPI alarms and the causes of network faults previously performed for our technical optimisation and clustering analysis.

Our algorithm employs a tree structure to extract features for the KPIs inputs of CSSR, DCR, HOF and TR. The selected KPIs are the dependent variables that are included in our data model shown in table 5.1. The rules based on KPI values will support the classification of the cost classes. Our datasets contain raw data and derived variables from our previous experiments.

TABLE 5.2 INPUT TRAINING DATA Input Variables (X) Key Performance Indicators (X) KPI Alarms (X) Class Labels (Y) Period

(Data/Time) (Call Success CSSR Rate)

NORM Maintenance Cost

BSC

Base Station Controller Location

TM

(Traffic Model) CRITICAL Infrastructure Cost

Standalone dedicated control channel HOFR (Handover Failure Rate) WARN Transmission Cost SDCCHAvailabilityRate SDCCHCR SDCCHAccessRate SDCCHDropsExcessiveTA SDCCHDropSuddLostCon Traffic Channel CDR (Call Dropped rate) Traffic Cost TCHTR TCH Seizure Attempts TCHCR TCHSS TCH Drop TCH Drop Rate TCH Availability Handover

HOFR (Handover Failure Rate) HOSR (Handover Success Rate)

Cost optimisation algorithm:

Aim: evaluate the network optimisation costs and predict the network areas that emerge to be optimised.

1. Purpose: Generate a cost classification based on cost evaluations of KPI metrics and KPI alarm values given by the optimisation algorithm in 4.5.1

Input data: The input data used for this experiment are shown in table 5.1. Training

data set with 2100 cell data referring to 150 locations in the area of Dublin in Ireland. Dependent variable: optimisation cost split into infrastructure costs, traffic cost, transmission costs and maintenance cost, a multi-class variable which is defined based on the KPI metrics and fault alarms values. Selected KPIs that used as an input of the cost optimisation algorithm referring to a certain amount of base stations (i=n)

are Dropped Call Rate, Call Success Rate, Traffic Rate and Handover Failures {(DCRi,

Rule 1: Transmission costs

The cells of several base station locations that are chosen to be optimised accordingly to the transmission cost rules are the ones that they present a high Dropped Call Rate and simultaneously the Call Success Rate is lower than usual even if the traffic rate is very low. In conclusion, evidence that show that in a not so busy network (low traffic rate) the dropped call rate is unusually high and the call success rate is lower than expected, transmission equipment need to be integrated. Hence, the rules defining the transmission costs for the model are as follows:

If DCR>2 and CSSR<=95% while TR<70% then

classify the cells in the area location as “Transmission Cost” Rule 2: Traffic cost

When the Traffic network rate is relatively high, the call success rate is low and the Handover Failures are greater than usual then the traffic hardware equipment may be checked for further upgrades. The rules defining the transmission costs are as follows:

if TR>70% and CSS<90% while HOF>=10% then

Rule 3: Infrastructure Cost

According to our previous experiment an obvious evidence of infrastructure issues are the KPI alarms that are resulting as “WARN”. So the cell areas that present

a high Handover Failures rate coupled with a “Warn” status given by the fault

optimization algorithm have to revise the hardware configuration. The rules defining

the Infrastructure costs are shown below:

else if KPI Alarm = “WARN” given by the fault optimisation algorithm while HOF >=10%

then classify the cells in the area location () as “Infrastructure Cost”

Rule 4: Maintenance Cost

Maintenance actions are required for all the cells that are categorsed as ”NORM” from the fault optimisation algorithm. These costs include network upgrades, software configurations and personnel costs. The rules for estimating the maintenance costs are as follows:

do if KPI Alarm = “Norm” class given by the fault optimisation algorithm then classify the cells in the area location () as “Maintenance Cost”

Output 1:

Applying the C45 algorithm by removing the input attributes of your model to the cost classes to construct new decision tree rules using the subset of training cell data for the selected locations. The purpose of the decision tree algorithm is integrating cost and error rate in decision tree pruning to find hidden relations between the data and KPIs that are not part of the input variables.

Output 2:

A cost sensitive classification model based on Naïve Bayes network algorithm to show the cost probability of the attributes of data and KPIs of the training set. Naive Bayes classification has mainly examined how to reduce the misclassification costs by reviewing different classification risks based on our algorithm rules (Chai et al. 2004).

For classification, the Naive Bayes algorithm computes the posterior probability of the data and KPI alarms belonging to the cost class defined according to Bayes’ rule. When missing values exist in a sample with the cells in different locations, the corresponding (output) attributes are simply left out in likelihood computation and the posterior cost probability is computed only based on the known (input) attributes (Chai et al. 2004). It is important to consider that both the output results are referring to the attributes of the data set that are not included in the selected input variables.

The algorithm employs a tree structure to extract particular features for particular inputs that refer to it as the Cost-Sensitive Tree of Classifiers (CSTC). Initially, the foundational concepts regarding the CSTC tree has been introduced that justify the derivation of a global cost term that extends to trees of classifiers. Next, the resulting loss function drive the model into a well-behaved optimisation problem (Xu, et al. 2014). The optimisation algorithm has also the ability to optimise the cost parameters based on the posterior probabilities that are determined by the neural network for the purpose of performance maximization.