Fig. 4. - GA Startle Process Conceptualisation
 
TABLE I
- FUZZY LINGUISTIC RATINGS
 
Linguistic Rating Terminology(Judgement of Influence)
Triangular Fuzzy Numbers(Numerical rating of factor’s influence)
Very Low Influence
0, 0, 0.25
Low Influence
0, 0.25, 0.50
Medium
0.25, 0.50, 0.75
High Influence
0.50, 0.75, 1.00
Very High Influence
0.75, 1.00, 1.00
 
TABLE II
- HUMAN CAUSAL FACTORS FROM THE HFACS RANKED BY AGGREGATED VALUES AS DETERMINED BY A PANEL OF AEROSPACE AND AVIATION
EXPERTS
 
Concepts
Causal Factors (Independent Variables)
LC
AR
MK
TH
AH
RM
SP
JS
Ranking
C1
Insufficient Training/ Lack of Concurrency
0.75
0.75
0.75
0.75
1.00
1.00
1.00
0.75
0.84
C2
Unskilled Pilot (Not rated for Aircraft Type for instance)
0.75
1.00
0.75
0.75
0.75
1.00
1.00
0.50
0.81
C3
Fatigue/Tiredness
0.50
0.75
0.50
1.00
1.00
1.00
0.75
0.75
0.78
C4
Faulty/Uncalibrated Instrument Readings
1.00
0.75
0.75
1.00
1.00
1.00
0.25
0.25
0.75
C5
Appraisal of Evolving Situation
0.75
1.00
0.75
1.00
1.00
0.50
0.50
0.50
0.75
C6
Medication/Drugs
1.00
0.75
0.50
0.25
1.00
1.00
0.50
1.00
0.75
C7
Communication (ATC)
1.00
0.75
1.00
0.50
1.00
0.50
0.25
0.25
0.66
C8
Stress
0.50
0.75
0.75
0.50
0.75
0.75
0.75
0.50
0.66
C9
Availability of Visual References
0.75
0.75
0.50
0.75
0.75
0.75
0.25
0.50
0.63
C10
Preparation (Flight/Route Planning, Pre-Flight checks etc)
1.00
0.25
0.50
1.00
1.00
0.25
0.25
0.25
0.56
C11
Resource Awareness/Crew Resource Management (CRM)
0.25
0.50
0.25
0.75
1.00
0.75
0.25
0.75
0.56
C12
Lack of ADM knowledge (Perceive – Process – Perform)
0.25
0.50
0.25
0.50
1.00
0.50
0.75
0.75
0.56
C13
Distractions (Phone Call, In Flight Conversations)
0.75
0.50
0.5
0.25
1.00
0.50
0.5
0.25
0.53
C14
Cockpit Ergonomics/Information Layout
0.50
0.50
0.25
0.75
1.00
0.25
0.25
0.75
0.53
C15
Time Pressures
0.50
0.50
0.25
0.25
1.00
0.50
0.50
0.75
0.53
C16
Lack of Assertiveness
0.50
0.75
0.25
0.25
1.00
0.50
0.25
0.25
0.47
C17
Complacency (Route Familiarity)
0.75
0.50
0.00
0.25
1.00
0.50
0.25
0.00
0.41
C18
Norms
0.25
0.50
0.25
0.25
1.00
0.25
0.25
0.50
0.41
C19
Part 91 Rules (Less Stringent Rules)
0.50
0.00
0.25
0.00
0.50
0.25
0.50
0.00
0.25
 
TABLE III
- SUMMARY OF EXPERTS WHO CONTRIBUTED TO THE RANKING OF CAUSAL FACTORS AGGREGATED FROM HFACS
Expert Initial
Occupation
LC
Chief Engineer (Aerospace Safety Systems)
AR
Aerospace Design Engineer
MK
Aerospace Design Engineer
TH
Aerospace Manufacturing Engineer
AH
Ex-CAA Safety Expert
RM_PPL1
Aerospace Engineer & GA Pilot
SP_PPL2
Aerospace Engineer & GA Pilot
JS_PPL3
Aerospace Engineer & GA Pilot
 
 
 
 
 
 
 
 
TABLE IV
- STARTLE DRIVER CONCEPT SUB GROUPINGS AND ORDERED RANKINGS ALIGNED TO HFACS GROUPINGS
 
HFACS CONCEPT GROUPINGS FOR FCM
Concepts of Acts & Omissions
Concept Description
Rating
C6
Medication/Drugs
0.75
C10
Preparation (Pre-Flight Checks)
0.56
C11
Awareness (CRM)
0.56
C16
Lack of Assertiveness
0.47
C17
Complacency
0.41
Concepts of Preconditions & Local Factors
Concept Descriptions
Rating
C2
Unskilled Pilot
0.81
C4
Faulty/Uncalibrated Instruments
0.75
C9
Visual References
0.63
C15
Time Pressures
0.53
C14
Cockpit Ergonomics (Information Layout)
0.53
C13
Distraction (Inflight)
0.53
Concepts of Supervision & Local Management
Concept Descriptions
Rating
C5
Poor Situation Appraisal
0.75
C7
Poor Communication (ATC)
0.66
C12
Lack of ADM Knowledge/Training
0.56
Concepts of Organisational Influences
Concept Descriptions
Rating
C1
Insufficient Training
0.84
C3
Fatigue/Tiredness
0.78
C8
Stress
0.66
C18
Norms (Familiarity)
0.41
C19
Part 91 Rules
0.25
 
TABLE V
- INITIAL MAPPING OF  THE  FCM MODEL
Initial Concept Mapping
Description
1 ßà 12
Insufficient Training -- Lack of ADM
1 ßà 5
Insufficient Training -- Poor Situation Appraisal
1 ßà2
Insufficient Training -- Unskilled Pilot
6 ßà5
Medication/Drugs -- Poor Situation Appraisal
6 ßà7
Medication/Drugs -- Poor Communication (ATC)
6 ßà16
Medication/Drugs -- Lack of Assertiveness
2 ßà10
Unskilled Pilot -- Poor Preparation (Pre-Flight Checks)
2 ßà16
Unskilled Pilot -- Lack of Assertiveness
2 ßà11
Unskilled Pilot -- CRM Awareness
2 ßà5
Unskilled Pilot Poor Situation Appraisal
5 ßà 16
Poor Situation Appraisal -- Lack of Assertiveness
15 ßà 8
Time Pressures -- Stress
9 ßà 5
Visual References -- Poor Situation Appraisal
4 ßà12
Faulty/Uncalibrated Instruments -- Lack of ADM
13 ßà 7
Distractions -- Poor Communication
3 ßà5
Fatigue/Tiredness -- Poor Situation Appraisal
14 ßà5
Cockpit Ergonomics -- Poor Situation Appraisal
14 ßà7
Cockpit Ergonomics -- Poor Communication
 
Thus,   a final vector Af is obtained, where the decision concepts are assessed to clarify the final decision of the specific decision support system. Essentially, the network automatically finds any relationships that may exist in the input data and then translates any discovered relationships into outputs – a form  of unsupervised learning where there is no training set of data implying no feedback from the network environment/system
Modification of the weight matrix of the mapped concepts for what-if analysis is made possible using tried and tested learning algorithms [41], [51], [54], [57], [59]. According to [57], [60], three main approaches for handling the task of FCM training emerge; These are Hebbian (which may be either signal, competitive, differential or differential competitive), evolutionary and a  hybrid (of the two previously mentioned) type of machine learning algorithms. The coverage of these algorithms is extensive in the referenced literature but are outside the scope of this paper. Practically, the mechanism of network state updates into the mapping matrix is done at each time step. This is achieved by using a modification of the current state vector’s sequencing, in a manner such that the value of wij, and the edge linking concepts Ci, and Cj are provided by a discrete version of the differential Hebbian law. The activation Hebbian learning (AHL) process, which this represents, essentially provides a procedure where the weight matrix of the FCM through time steps is modified to model the system’s behaviour iteratively. Mathematically, this discrete version takes the form:
Wi j (t + 1) = wi j (t) + µi (∆Ci (t). ∆Cj (t) – wi j (t)   (7)
Where ∆Ci is the change in the ith concept and
∆Ci (t) = Ci (t) – Cj (t – 1)          (8)
The learning coefficient µt decreases gradually over time, based on the following equation:
µt = 0.1[1 – t/1. 1N]    (9)
with the constant N to ensure µt remains positive. Viewing the fact that there is no consideration of the time relationship between the concepts, the model can be thought of as a general representation of the scenario or system being modelled. These heuristic methods facilitate a good estimation of near-optimal solutions, with a pragmatic optimisation of the error function. In the FCM Expert tool, concepts in the map are activated by making their vector element fall within the range [0, 1]. The threshold function mentioned earlier, reduces the boundless weighted sum to a predetermined range, facilitating a qualitative comparison between and across concepts, thus, representing the fuzzy linguistic associations in the graph. In literature, there are three main threshold functions, namely the Bivalent, Trivalent and the Logistic Signal, a special case known as the Sigmoid function [33]. The Sigmoid (bipolar) function is chosen for this present work because it seems to offer significant advantages over the others, especially where vision system performances and eye tracking are concerned [58], [61], [16]. The Bivalent and Trivalent options are considered restrictive for the aim of this study and are not considered.
 
The following graphs (Figures 5 – 7) illustrate the efficacy of the Fuzzy Cognitive Mapping process using the FCM Expert tool, highlighting different “what-if” scenarios considered during the mapping process, based on input from experts. Effectively, the outputs provide an objective ranking of the causal factors in terms of their propensity to cause a pilot to startle, as a root cause. For the efficacy of analysis and the benefit of progress, the top four concepts (Table IX) below, following the mapping process, are considered for aligning the embodiment of future experimentation into startle impacts on performance.
In the map of Figure 5, for example, Startle itself is not assigned the role of a decision concept, but instead, as a concept receiving inputs from the other causal concepts.Further interaction amongst concepts is captured in subsequent iterations as per Table V and Figure 1 earlier. Figure 5, in particular yields an output (a hierarchy of causal factors) determined by the FCM through an inference algorithm based on population heuristic search methods. In this first case (Figure 5), the final outputs converge to a top-four causal factor hierarchy of Concepts 5, 9, 7 and 2 – in that order, as most critical to the cause of startle. From Table II, these distil to Poor appraisal of the situation; Poor visual references; Poor communication skills (especially with air traffic control); Unskilled Pilot. These are all plausible factors in the case of novice GA pilots (of interest to current research), but can also be the case for even more experienced pilots. This demonstrates the power of the FCM to extract causal factors for an intangible output concept such as startle.
Figure 6 is an example of the iteration test on the initial mapping created. In this case, as in all other iterations tested, Startle is still not a  decision concept and is considered in the context of other interactions amongst concepts.  This map was tuned to test what impact adjusting the fixed-point attractor would have. Setting epsilon here to 0.01, instead of 0.001. Concepts C5 (Poor situation appraisal); C1 (Insufficient training); C7 (Poor Communication with ATC) and C2 (Unskilled Pilot), are shown to be the top four dominant factors in this scenario. Again, these outputs are like the first iteration except for the top concept – a lack of training. These outputs, as have been discussed, highlight how the FCM makes it easy to draw the connection between these concepts and poor performance, should a startling event occur. Figure 7 exemplifies the more complexity mapping – Similar to Case B but with return connection between concepts as highlighted in Table V. Startle, in this case, is also considered static. In this example, C16 (Lack of assertiveness); C12 (Poor ADM Knowledge); C10 (Poor preparation) and C8 (Stress) – See Table VIII (Item LB-A-008) present themselves as the top four driving factors of startle. Once again, it demonstrates the capability of the FCM method to remove subjectivity in the process of understanding the causal mechanisms of startle.
The top four concepts post-convergence are deemed to be appropriate for material embodiment in experimentation protocols in a manageable way. Tables 6 - 8 are supplied to demonstrate how the mapped models provide inferences on the connectedness of concepts driving the output of startle. They also provide the reader with a visual assessment of the general behaviour of the models. Outputs of Table VIII in particular present a contextual basis for the analysis in this research. This is because it provides outputs of the randomised weighted input maps. A basic modal analysis strategy then provided the best choice options from the iterations, yields (Table IX), the auto-randomisation of concept weights through iterations, adds a layer of robustness and a high level of objectivity since the author/experts had no input to the initialised weightings.
 
System impact considerations for this work is driven by outputs of the previously discussed FCM process involving abstracting interrelationships of causal elements. In the preliminary stages of formulating an appropriate methodology, a series of flight training tasks were completed in a flight simulator [62] by the researcher and a group of students. These tests were intended to establish a qualitative basis upon which startling test cases can be created in the simulator space to capture the outputs of the FCM (clear-air turbulence scenario was selected - ). It was then possible to develop structured and logical consideration of startle impact on performance, through design, simulations, visualisations of gaze capture and the evaluation of the complex human-machine interaction. Such strategies have been used to successfully model and evaluate processes characteristic of human interaction with complex systems to great effect [52], [53], [63]. The Salience, effort, expectancy and value (SEEV) framework [24], [64], plays a central role in guiding this research. The reason for this is that it allows linking of a human factors model and a computational cognitive structure, to represent human capabilities and limitations. [29], [49], [64].
As gleaned from the outputs of questionnaires, it is then possible to chart the potential routes to impact on performance, using the FCM of human factor concepts. The concepts used in the mapping are chosen based on inferring the predominance of visual perception, visual attention, task management, decision making as well as memory mechanisms abstracted from the Human Factors Analysis and Classification System and aligned with the SEEV framework. Figure 8 depicts the Structure of the SEEV system. This structure is based on the foundational work by [20], [65] on the Aviate, Separate, Navigate, Communicate task execution process, and represents an optimal expectancy model as it pertains to expert pilots. This structure is considered useful for inferential analyses, in the sense that it fosters a reflection on the “what-if” scenario to provide an insight into the startle effect modality and impact on the task performance of a novice pilot where attention bandwidth is reduced.
 
 
 
 

V.    SUMMARY

Three key agents were resolved as being critical to the development of an understanding of startle impact represented in physiological outputs and task performances. These include the Human Mental Model (reduced visual conditions and task complexity), Aircraft Mental Model (Aerodynamic and Stability Behaviour as a function of pilot inputs) and the Environment Dynamics (The physical world represented in the Simulator space). Therefore, a modular hybrid approach would provide the best opportunity for developing a coherent view of startle causality. Much like one proposed by [66], which includes two fundamental approaches; an equation based mixed system model with constituent elements (agents) of the system clearly defined, complemented by an embodiment of the elements, into a hardware-software architecture for experimentation. Thus, the articulation of capability requirements is assured to train a novice pilot and testing startle impacts. In order to embody the first part of such an approach, in this case, Fuzzy cognitive mapping was implemented.
The FCM evidence, discussed in this paper shows its usefulness for the current research within four key functions: explanatory, predictive, reflective and strategic. In the explanatory context, the FCM provides a flexible and robust means to reason out an objective representation and therefore, understanding of causality, while highlighting any distortions and limits of the use case representation. The predictive function involves, as the name suggests a prediction of future decisions, actions and tendencies which an agent would give to justify any new instances of a concept, as the starting weight of each concept as iterated. In this case, we can investigate the concept which presents the most risk output by the FCM convergence plots. This prediction function lends itself to modelling, and the subsequent analysis of any experiment results collated from the context of the mapped outputs. The reflective function of an FCM provides a means of judging the adequacy of a decision profile.
In contrast, the strategic function can be used to generate a more accurate description of a complex scenario representing a highly dynamic psychophysiological response of an operator. As was shown in Table IX, the auto-generation of weight initialisation within the map for eight iterations diluted the problem space to four critical concepts as the drivers of a startled reaction. These have been determined as Concept 2 (Unskilled Pilot); Concept 1 (Insufficient training); Concept 16 (A lack of assertiveness), Concept 6 (Medication/Drugs)  in that order. Interestingly, given the randomisation of the iterations as Table IX suggests, these top four items sit well as a plausible root cause for a pilot to be startled given an abrupt activation of inactive frames during L1 SA and subsequently actioned by a fast appraisal of the situation. The contention here is that such fast appraisal at the lower levels of situational awareness can lead to an instinctive but wrong application of control laws. Incidentally, as this research seeks to understand the implication of being startled for the novice/fledgeling GA pilot, this bodes well for guiding further experimentation even for experienced pilots in general. At the very least, assurances can be extracted from the mapping process that if the pilot is unskilled and has insufficient training (i.e. novice), then they would be far less assertive given an evolving situation, which may ultimately cause them to startle.
Furthermore, should they have some level of training but are hampered by prescription medication, for instance, then, this would also very likely make them startle easily. Given the stringent rules for piloting an aircraft, this fourth item of causality (drugs) is thought to be highly unlikely but cannot be ruled out from human factors. Given the findings from the mapping, experimental studies can then be objectively defined in more detail, such as is fit for understanding the dynamic process and execution of the startling input to normal operations. It is expected that the experimentation and subsequent interpretation of any results thereof, will find alignment with the hierarchy of causal factors as have been determined and should help guide the development of protocols which can be  positively transferred to real life operations.