The model developed gives the probability of the survival of the aircraft, PS, depending on the states of variables representing the surrounding world, knowl-edge about an emerging threat, and a selection of possible actions for the pilot to take.
The first step is to specify the structure of the BNbased on knowledge about the EW domain. This structure can have several layouts, depending on the degree of details one wishes to model. The general rule is to incorporate enough details to be able to establish the probability distribution between nodes. On the other hand, more details will require more nodes, and thus populating the dependency tables becomes more difficult. When the structure of the BN is in place, every node is supplied with a sufficient number of states, and the conditional probabilities between nodes are entered into the dependency tables.
At first the model is a ”pure”BNsince no decision or utility nodes are introduced.
Any probabilities used in the model are merely based on ”good guesses” and not on data from real aircraft and threats. Adding decision and utility nodes to theBNwill change it into a DG. As explained in Section 5.2.7 no good utility scale is found for the use of aDGin decision support for fighter pilot, and hence no utility nodes are used in the model.
5.3.1 Assumptions
It is assumed that the current combat scenario contains at most one threat, and that only a single missile may be launched towards the aircraft at any given moment. This will ensure that when multiple nodes receive evidence they are all the result of the presence of a single threat. If more threats are present in the scenario, and they are detected by e.g. radiation in severalRFbands, the probability of a proper detection of each of these by theBNdecreases.
In the Prolog approach (Chapter 4) it is assumed that all warnings relate to real threats or friendly aircraft. In theBNapproach this is not the case. Here it is assumed that all observations come with a probability/certainty, and thus a warning indicates that a threat is present with a given probability. Since none of the on-board sensors/warners give this number in conjunction with a warning it is up to athreat evaluator to calculate it. This can be done using statistics on the number of times a warning was given for a threat in a given distance and at a given angle, etc. This threat evaluator is assumed existing and working, and the design of it is not considered part of this work.
5.3 Building the Model 87
Figure 5.9: TheBNmodelling the world in and around the fighter aircraft. The model is divided into three layers, one representing the world surrounding the aircraft, another representing the on-board sensors, and the last one giving the survivability for a given combination of states having received evidence. (Picture exported from HUGIN.)
5.3.2 Constructing the Model
The BN constructed can be seen in Figure 5.9. It is lay out as a three layer model, with the top layer representing the world surrounding the aircraft, the middle layer containing nodes for the on-board sensors, and finally the bottom layer giving the results of deploying countermeasures. The jammer is placed in the middle layer, as it gives input about RF locks. Since it serves as both a sensor and a countermeasure it might as well have been placed in the bottom layer.
All edges between the ordinary nodes in the BNshows causality. To ease the reading of the model the nodes are arranged in such a way that causality points downwards. In this way the nodes in the top layer (the surrounding world), causes changes in the states of the nodes in the middle layer (the sensors), which in turn influence the calculated survivability.
88 The Bayesian Network Approach
In the construction of the model some observations are made. A number of these are described below.
Enemy Territory. If enemies are present and threatening the aircraft, the aircraft is, by definition, flying over enemy territory. This is why the Enemy Territory is not just a binary node, indicating whether or not the border has been crossed; enemies might exist on the friendly side of a border as well (e.g. terrorists), and missiles may be ”friendly fire”, i.e.
launched by ones own forces.
Missile Seen. It is possible for the pilot to visually see an incoming missile that he has not been warned about by either the MWS or the RWR. To represent this, a Missile Seen node was initially added to the BNmodel.
While the presence of a missile that has not been seen by the aircraft sensors is of vital importance to the results of the DSS, it may not be possible to use this information in the DSS. To do so would require the pilot to tell the system that a missile has been seen, and possibly both the direction and range to the missile. While it is possible to construct and incorporate this type of registration in thePVI, the registering process in itself would be too time consuming to be feasible. The missile may have hit the aircraft before the pilot has told the system about it. For this reason the node is not part of the model.
Expendables. In the model dispensing chaff or flares is described using a bi-nary action node; either they are dispensed or they are not. In the real world there would be more programs to select from, each program de-signed to take care of a given threat or threat scenario. Introducing more programs into the model will increment the number of entries in the de-pendency tables, which again will make it more difficult to find proper values for making theBNbehave properly.
Guidance System. It is assumed that any missile guidance system uses either
RForIRguidance, and that none of the missiles under consideration in this work has multiple guidance systems. Therefore the detection of a radar guided missile, as seen by theRWRcan intuitively lead to the assumption that a missile detected by the MWS, assuming it is detected in the same direction, will not be using IRguidance. This gives the causality between two nodes, RF GuidanceandIR Guidance, as seen in Figure 5.10.
RF Guidance IR Guidance
Figure 5.10: If a missile isRFguided it will not also beIRguided.
5.3 Building the Model 89
SinceRFandIRguidance are mutual exclusive there is no reason to main-tain two nodes, and the RF Guidance and IR Guidance nodes are thus merged to a single node namedGuidance.
Since different types of missiles use the same type of guidance, with dif-ferent results, the presence of an IR lock depends on both the type of missile and on the guidance in use. It can not depend on the missile type alone, since missiles having the same missile types may be equipped with different types of guidance systems.
Decision and Utility Nodes A number of decision nodes are added to the model. These nodes and their states can be seen in Table 5.4. Let |A|
be the number of states in a decision node A. With the seven nodes in the model there are |Jammer Present| · |Jammer Mode| · |Manoeuvre| ·
|Chaff Loaded| · |Chaff| · |Flares Loaded| · |Flares|= 2·3·3·2·2·2·2 = 288 combinations of states in the decision nodes.
Not all decision nodes need to be part of the decision since the pilot has no way of changing the state of these while in-flight. The decision nodes is thus split into two sets, the preparation nodes comprising the Jammer Present, Chaff Loaded, and Flares Loaded, and the action nodes:
Jammer Mode, Manoeuvre, Chaff, and Flares. This now gives a total of 3·3·2·2 = 36 combinations of states in the action nodes which should be tested to find the combination yielding the highest survivability. Some of these combinations would not be feasible, and could thus be removed from the set of combinations to test for. For instance a combination of flares and manoeuvres does not make any sense, if Flares Loadedindicate that no flares are available.
Table 5.4: The decision nodes and their states.
As can be seen in Figure 5.9 the action nodes are interconnected. In Section 5.2.6 it is stated that action nodes have no dependency tables, and the edges between action nodes does not constitute a parent-child relation per se. Instead they are needed for the propagation algorithm used by HUGIN. The direction of the arrows between the action nodes is of no influence to the calculated survivability.
90 The Bayesian Network Approach
As described in Section 5.2.6 the introduction of a utility scale for the model is not straightforward. At first the number of remaining expend-ables was introduced as a utility scale and utility nodes were connected to the Flare and Chaff nodes. This is skipped since flares and chaff are two very different types of expendables, and having a number of flares left will not increase the survivability when a RFbased threat occurs. Sec-ond, optimizing the amount of inventory would only result in using less expendable, not in increasing the survivability at all. For these reasons no utility nodes are used in the final model.