Information Technology Reference
In-Depth Information
two subsets A1 and A2, both of length 10. A1 represents the training data set
from which the TEA will develop a long term memory of trends associated with
A1. A2 represents the testing data set which the TEA will have to examine in
the light of information preserved from the experience of A1.
A1 contains three simple trends, T1, T2 and T3. They are closely related,
in terms of the price movements they contain, so presenting A1 to the TEA
represents a simple challenge to ensure the TEA operates correctly.
A2's purpose is to test the ability of the TEA to handle a more complex
antigen with more diverse trends. A2 comprises 6 trends, T1 and T3 as were
noted in A1, in addition to four new trends T4 to T7. Compared to A1 we have
increased the number of trends from three to six and increased their length and
diversity, making it more dicult for the TEA to find all the trends in A2.
It is hypothesized that although trends T4, T5, T6 and T7 are more com-
plex to identify from knowledge of A2 alone, after experiencing trends T1, T2,
and T3 from A1's presentation, which are related to T4 to T7, the TEA should
develop some form of association between the trends leading to an easier recog-
nition of these new patterns. To test this hypothesis we define the following 4
experiments.
In experiment 1 the training set A1 will be presented to the TEA from gener-
ations 1 to 10. The testing set A2 is then presented to the TEA from generations
30 to 40. The TEA is run for 50 generations and the experiment repeated and
results averaged across 10 runs. The frequency of detection of trends T1 to T7 is
recorded across all runs. To give a base line comparison where there is no mem-
ory in the system experiment 1 assumes no knowledge of A1 is carried forward in
the TEA during A2's presentation. At the point of A2's presentation the tracker
population is replaced by the random tracker population created in generation
0. The TEA therefore has to learn to recognise trends in A2 from scratch.
Experiment 2 investigates the impact of incorporating feedback from the long
term memory pool into the TEA. We repeat the previous experiment, but the
tracker population at generation 30 is repopulated using clones from the long
term memory pool. We identify whether any association properties become ap-
parent in the TEA by examining the frequency with which the trackers in
the long term memory pool have detected the trends in A2 as compared to
experiment 1.
Experiment 3 investigates the issue of scalability in the TEA. Experiments 1
and 2 present antigen sub sets of only 10 data items at a time. We now scale
up the information presented to evaluate the impact on the TEA's performance.
Experiment 3 presents the complete antigen A to the TEA from generation 1 to
20, doubling the size of the information presented. Results in terms of population
sizes, trend detection rates and memory pool eciency are then to be compared
with experiments 1 and 2.
Experiment 4 compares the performance of the TEA against a random search.
Each tracker generated during execution represents a potential search solution;
given the high population levels anticipated in the TEA one could argue that a
large randomly generated tracker population would also succeed in identifying
