Exploratory data analysis for airline disruption management

Artificial Neural Network Modeling for Airline Disruption Management

“…For instance, a single hidden layer perceptron provides an excellent medium for solving linearly separable problems and can also be used to define a Gaussian process [18][19][20]. As such, we define the topology of our artificial neural network as a single hidden layer perceptron, based upon the following properties gleaned from outcomes of previous work on exploratory data analysis for ADM [21]:…”

“…The ANN framework, shown in Fig. 3, represents a feedforward perceptron that consists of an input layer 𝑎 with a set of aleatoric and epistemic flight schedule features [21] or nodes 𝑓 𝑎 , a hidden layer 𝑏 that contains nodes 𝑓 𝑏 which transmute information from nodes 𝑓 𝑎 in the input layer, and an output layer 𝑐 with a flight schedule feature or node 𝑓 𝑐 that represents performance measures at any stage of flight schedule execution. For validation of the ANN topology, predicted values of performance measures are compared with actual (realized) values to inform the adjustment of the ANN parameters for improving its predictions.…”

“…5 shows our solution approach for automatic PTFM estimation of performance measures for ADM. The approach begins by preprocessing raw data into readily decipherable data formats that are suitable for ANN learning algorithms, by employing the abstraction and techniques described in [21]. Next, we learn the optimal network topology of each ANN for predicting target flight schedule features at each stage of airline resource capacity management, before performance measures at each stage are estimated through a combination of flight schedule features inferred from the optimal ANNs.…”

“…Table 1 summarizes specific epistemic input flight schedule features and corresponding target features in all ANNs for separate phases of ADM. These features represent inputs and targets that are applicable for separate classes of ANNs for each phase of ADM, and are adopted based upon the results observed from exploratory data analysis [21]. Furthermore, all determinate aleatoric flight schedule features represent additional generic inputs for all ANNs for each phase of ADM.…”

“…shows the summary of the model performance of separate ANNs in estimating appropriate flight schedule features for all functional roles[21] in the AOCC during disruption management. As shown in Table2, all ANNs for operational ADM have near perfect performance with AUC values of approximately 1.…”