Information Technology Reference
In-Depth Information
is whether the TEA can learn, through feedback from its long term memory
pool, to associate what it has memorised from previous experiences to aid in
the investigation of new novel antigen. Comparing the results of experiments 1
and 2 we see incorporation of the memory pool has a beneficial effect on the
ability of the TEA to map to and memorise trends in a more complex antigen.
Compared to its naive counterpart the inclusion of the long term memory pool
boosts trend recognition from 55.7% to a potential 87.%, whilst ineciency in
the memory pool is kept consistently low at 2.1%.
The reason for this improvement can be seen if we view the trends within
the antigen subsets A1 and A2, as shown in Table 1. Trends T1, T2 and T3,
located within A1, have price change combinations involving rises of $1 or $2.
Recognition and development of memory trackers associated with these trends
would assist the TEA in identifying trends T4, T6 and T7 in A2 as they too
have price combinations that involve price rises of $1 and $2. If memory trackers
can be successfully evolved to map to these trends during presentation of A1,
as was shown in experiment 2, then the TEA can utilise that knowledge and
associate new novel trends with those already seen, instigating a more successful
response. Without the ability to associate new experiences with past knowledge
the performance of the TEA declines significantly, as expected.
Although the antigen investigated here is very small and simplistic, it is im-
portant for the TEA to scale to handle larger antigens. Experiment 3 gives us an
indication of the scalability of the system as antigen sizes increase. Comparing
test experiments 2 and 3 we see increasing the antigen size by 100% from 10
to 20 causes the maximum tracker population to increase from 323 trackers to
2,244, leading to an exponetial growth problem. Splitting antigen A into it's two
component parts, as done in experiment 3, results in significantly lower popula-
tion sizes whilst still maintaining a high detection rate. This is only possible if
we carry forward the long term memory pool and feed it back into the tracker
population to assist in future antigen recognitions. In this way we can avoid the
scalability issue whilst maintaining a high degree of accuracy in the TEA.
However, from test 3 it was evident that separating antigen A at the mid point
results in trend T8 now not being recognised as a trend within the component
parts A1 and A2. T8 exists within A1 and A2 but is not repeatedly stimulated so
has a SF of 1, therefore it does not conform to the definition of a trend in either
A1 or A2. To avoid this issue the algorithm could be re-run with alternative split
points to generate an overall memory pool; this will be investigated in future
work. From analysis in experiment 4 we can also conclude that the TEA performs
significantly better than a random search in identifying trends prevailing in a
small data set.
8Conluon
This paper presents an immune inspired algorithm that is successful in identify-
ing trends in a small simple data set. The authors theorise that these techniques
can be expanded and applied to larger time series data sets to identify trends
