The information collected within this review can now be used to frame the research questions presented in Chapter 1 within the context of the current state of the art within the fields of artificial immune systems and robotics.
2.5.1
Migrating from Software to the Physical World
Can an immune-inspired, anomaly detection algorithm be adapted to solve threat detection problems in the physical world, through the medium of a robot?
The current state-of-the-art in security robotics has several key limita- tions. Most importantly, all of the systems encountered were unable to tolerise against dynamic objects, constraining them to being used in empty warehouses or outside the hours of operation for a building. Sec- ondly, the inability of the robots to perform threat detection while mov- ing severely limits their applicability in the real world. In AIS computer security applications there has already been a recognition that ‘normal’ requires a more complex description. It would be totally impractical to have a computer security system that treated all network activity as a threat. As a result, there are already algorithms, such as negative se- lection and the DCA that attempt to differentiate between normal and anomalous behaviour in a more intelligent way.
2.5.2
Emergent Properties of the Dendritic Cell Al-
gorithm
Does the dendritic cell algorithm have properties that were not explicitly added as part of its design, which could be advantageous to a robotic application?
feature of robotic systems. As highlighted by the problems with other AIS techniques, it is important to ensure that abstractions of immune concepts are not simply ‘re-inventing the wheel’ so any investigation with the DCA at its core, must assess it from a theoretical stand point, not simply an empirical one.
2.5.3
Applying The Dendritic Cell Algorithm to a
Robot
Is it possible to adapt the dendritic cell algorithm from being a batch system to a system that can operate on a robotic platform?
As discussed, there is little work on theoretically analysing the DCA. However, there is precedent for alternatives to the batch system of gen- erating the MCAV. This suggests that the core functionality of the algo- rithm is independent of the cell analysis phase. When this investigation was begun, the implementations of the DCA were all performed within the ‘libtissue’ framework and very slow. It is an important step to guar- antee that the algorithm can be made light-weight enough to run on a robot on-line for the system to be a feasible solution. The work of Lay and Bate, while not directly applicable, raises important issues relating
2.5.4
The Benefits of the Dendritic Cell Algorithm
Are there other algorithms with functional equivalence to the dendritic cell algorithm, which can outperform it in terms of reduced computational complexity or superior performance, for the threat detection problem?
While weaknesses in the current solutions to the robotic security problem have been identified, a solution based on the DCA should be critically appraised. With so few results published in the literature, it is difficult to see how a direct like-for-like comparison could be produced. However, its performance should be at least as good as a random classifier, to demonstrate that information about the problem is being used to make appropriate decisions. In addition, it should be clarified if the DCA has functional equivalence to other techniques, or a collection of other techniques, and if such a system could outperform the algorithm.
“You see, but you do not observe. The distinction is clear.” - Sir Arthur Conan Doyle (Sherlock Holmes), A Scandal in Bohemia (1892)
3.1
Introduction
The work presented in this chapter is based on [107].
As discussed in Chapter 1, technologies and protocols designed to enforce security are now pervasive in society. Many houses now have burglar alarms and CCTV is common-place in towns and cities. Both Chapters 1 and 2 highlight the need for each robot within a robotic security system to be able to run as a self-contained agent. As a result any algorithm employed by a robotic security system must prove itself to be capable of running in its entirety on a mobile robot, without having a detrimental effect on core functions, such as obstacle avoidance.
At the time that this research was carried out the dendritic cell algo- rithm (DCA) had not been used as a real-time algorithm. In order for it to be useful as a processing algorithm for a mobile robot it is necessary to address architectural limitations that have previously constrained it to be a batch processing algorithm. Once engineered for use with a real-time processing system, the DCA must be interfaced to a robotic operating system. In addition to these modifications, the implementation of any version of the DCA requires the input heuristics to be specified and de- signed. Finally, it is necessary to parametrise the population distribution for the specific application. This is vitally important as it has been shown in [42] that the performance of the DCA is significantly effected by the correct identification of this distribution.
At the prototype stage it is important to verify that the performance of the DCA is of a sufficient quality to pursue. Later chapters will explore the DCA’s performance relative to other classification techniques using theoretical analysis. In this chapter the target shall be to outperform the calculated classification rate based on the expected results for a classifier operating at random. To outperform such a classifier is a minimum requirement of any potential algorithm.
an event or if the algorithm requires more computation time than the hard real-time constraint of 250ms per allocated processing slot. More information about these constraints is provided in Section 3.2.4.
• Untrained and using unfiltered sensor data, the DCA outperforms a random classification technique. This hypothesis can be accepted if a single parametrisation can be found which causes the average classification error for the DCA to be lower than a random classifier, theoretically calculated from the layout of the environment. In Section 3.2 the modifications required to implement the DCA on a real-time system are discussed. In Section 3.3 an experiment is de- signed to test the classification accuracy of the system. In Section 3.4 the preliminary results are presented and analysed and problems with the prototype are highlighted. In Section 3.5 a modification to the orig- inal system is presented that corrects one of the major short-comings of the initial system. In Section 3.6 the results of the improved system are presented and analysed. Finally, Section 3.7 discusses the contributions made by these experiments and work that has subsequently been carried out by other researchers.