Canon PowerShot SX610 HS Quick start guide

23
PC90
PC100
23
11.8
N=
INITIAL
23
23
23
23
PC10
PC20
PC30
PC40
ncurses
23
23
23
23
PC50
PC60
PC70
PC80
23
PC90
23
PC100
23
8.4
5.5
N=
INITIAL
23
23
PC10
PC20
23
23
PC30
PC40
nmh
2.4
8.6
7
23
23
PC50
23
PC60
23
23
PC70
PC80
23
PC90
23
PC100
23
N=
INITIAL
23
23
23
23
PC10
PC20
PC30
PC40
incl
2.3
23
23
23
23
PC50
PC60
PC70
PC80
22
2
1
1
9
18
8
N=
23
23
PC10
PC20
23
23
PC30
PC40
23
PC50
23
PC60
23
PC70
23
PC80
23
PC90
23
PC100
23
INITIAL
modssl
8.4
20
7
21
3
2.3
23
INITIAL
8.6
45.0
22
23
PC100
inn
45.5
22
9
23
PC90
1
44.5
8.2
7
12
20
9
8.0
8
12
21
2.2
2.2
23
15
12
15
23
1
44.0
3
43.5
7.8
43.0
16
1
15
7
2.1
20
3
13
14
2.0
23
23
23
23
PC10
PC20
PC30
PC40
23
23
23
23
PC50
PC60
PC70
PC80
23
PC90
23
PC100
mtunis
23
INITIAL
7.4
42.0
2.1
N=
7.6
42.5
8
21
2
41.5
N=
23
23
23
23
PC10
PC20
PC30
PC40
23
23
23
23
PC50
PC60
PC70
PC80
23
PC90
23
PC100
23
7.2
N=
INITIAL
23
PC10
23
PC20
23
23
PC30
PC40
23
PC50
23
PC60
23
PC70
23
PC80
23
PC90
23
PC100
23
INITIAL
swing
rcs
N=
23
23
23
23
PC10
PC20
PC30
PC40
23
23
23
23
PC50
PC60
PC70
PC80
23
PC90
23
PC100
23
INITIAL
xntp
Figure 4. Worst results obtained by using MDGs with weighted edges.
dition, weak results from the initial hill climb can also
achieve consistently better values (for example best and
worst performance for mtunis shown in Figures 1 and
2).
One reason for observing more substantial improvement in larger MDGs may be attributed to the nature
of the MQ fitness measure. Unfortunately the MQ fitness measure is not normalized, for example a double
increase of MQ does not signify a doubling of modularization quality. At best, we can only claim that MQ is
an ordinal metric [11]. To overcome this, the percentage MQ improvement of the final runs over the initial
runs is also measured (see Table 5). Using these values,
tests were carried out to determine any improvement correlating with the MDG complexity. The number of nodes
and the number of connections in each MDG were tested
for correlation against largest percentage improvement of
each of the final runs against the initial run. These statistical tests show no significant correlation between size
and improvement in fitness irrelevant of weighted or nonweighted MDGs.
Improvements are always achieved for selection cut
off values of 10% and 20%, in most cases there are improvements across all final hill climbs. However there are
exceptions. A dramatic example of this is in bunch (Figures 1 and 2), where results only show an increase for the
cases where 10% and 20% of the initial climbs are used
for building blocks.
5.3
of the algorithm were used. Furthermore this technique
was used on MDGs with weighted and without weighted
edges of different sizes to improve the strength of the results for more general cases.
Employing the Wilcoxen signed ranked test helped to
show that the improvements are significant enough to be
an unlikely chance occurrence. The reduction in variance caused by the selection mechanism may mislead
the Wilcoxen ranked test to find significant difference between the initial and final runs. Therefore actual improvement in the fitness over the initial runs were measured to
determine whether the search is capable of discovering
better peaks in the landscape.
6 Conclusions
The multiple hill climb technique proposed here has
produced improved results across all MDGs, weighted
and non-weighted. There is some evidence that the technique works better for larger MDGs but this could be due
to the ordinal nature of the MQ metric used to assess
modularisation quality.
This difficulty aside, larger MDGs tend to achieve relatively earlier benefits across the final hill climb runs
from this technique. For example MDGs with small number of nodes and edges tend to show little or no improvement until building blocks used for the final hill climb are
selected at 10% and 20%. On the other hand MDGs with
a large number of nodes and edges tend to show significant improvement on the initial search across most or all
of the final runs (Tables 2 and 3).
The increase in fitness, regardless of number of nodes
or edges, tends to be more apparent as the building blocks
are created from a smaller selection of individuals. This
Experimental concerns
Due to inherent randomness in any hill climbing
search technique, it is hard to identify any trends by looking at individual hill climbs. For this reason multiple runs
8
Proceedings of the International Conference on Software Maintenance (ICSM’03)
1063-6773/03 $17.00 © 2003 IEEE
PK
�ITL�&�z9.txtFigure 5. MQ increase against
number of edges for MDGs with
no weighted edges
Figure 6. MQ increase against
number Of nodes for MDGs with
no weighted edges
Figure 7. Percentage MQ in-
Figure 8. Percentage MQ in-
Figure 9. MQ increase against
number Of edges for MDGs with
weighted edges
Figure 10. MQ increase against
number Of nodes for MDGs with
weighted edges
crease against number Of nodes
for MDGs with no weighted
edges
crease against number Of edges
for MDGs with no weighted
edges
Figure 11. Percentage MQ in-
Figure 12. Percentage MQ in-
crease against number Of edges
for MDGs with weighted edges
crease against number Of nodes
for MDGs with weighted edges
9
Proceedings of the International Conference on Software Maintenance (ICSM’03)
1063-6773/03 $17.00 © 2003 IEEE
PK
�ITL�Z�4yy10.txtmay signify some degree of importance for the selection
process. Perhaps the less fit solutions in the initial population are more likely to represent the same peak in the
solution space and removing them by a more elite selection process may reduce the noise or bias this may introduce and increase the likelihood of a more concentrated
search.
This work is supported, in part, by EPSRC Grants
GR/R98938, GR/M58719, GR/M78083 and GR/R43150
and by the Brunel Research Initiative and Enterprise
Fund.
References
[1] C ONSTANTINE , L. L., AND YOURDON , E. Structured
Design. Prentice Hall, 1979.
7 Future work
[2] D OVAL , D., M ANCORIDIS , S., AND M ITCHELL , B. S.
Automatic clustering of software systems using a genetic
algorithm. In International Conference on Software Tools
and Engineering Practice (STEP’99) (Pittsburgh, PA, 30
August - 2 September 1999).
The selection techniques used in building block creation may be extended. This may well be achieved by
ensuring that all hill climbs in the initial stage are unique.
For very small MDGs this may cover all peaks in the
landscape, also some sections of the landscape may be
harder to search in the initial stage. Another possible
method to ensure improved selection would be to include
an attribute which determines the importance of each initial result in construction of the building blocks. This
attribute could be related to frequency or distribution of
the initial solutions.
In addition, other techniques to measure complexity of
the MDGs and use of a normalised fitness measure could
help in the recognition of any relationship between MDG
complexity and the improvement achieved by using this
technique.
Finally, once an improved selection technique is identified, multiple iterations of building block creation/ hill
climbs can be used to focus the search further. Alternatively this technique could be used to improve genetic algorithms (GAs). GAs have already been used for clustering but with generally worse results than pure hill climbing [2, 4, 8]. The less than ideal results might be a
consequence of the crossover operator in GAs, which is
deemed to be more effective when the structure of the
chromosomes aids the transmission of useful information
between generations [5]. The use of the building blocks
created using this technique to seed a GA could help to
preserve this information and improve the GA’s performance without resorting to complicated evolutionary repair operators [3].
[3] FALKENAUER , E. A new representation and operators for
genetic algorithms applied to grouping problems. Evolutionary Computation 2, 2 (1994), 123–144.
[4] H ARMAN , M., H IERONS , R., AND P ROCTOR , M. A new
representation and crossover operator for search-based optimization of software modularization. In GECCO 2002:
Proceedings of the Genetic and Evolutionary Computation Conference (New York, 9-13 July 2002), Morgan
Kaufmann Publishers, pp. 1351–1358.
[5] J ONES , T. Crossover, macromutation, and populationbased search. In Proceedings of the 6th International Conference on Genetic Algorithms (San Francisco, July 15–19
1995), L. J. Eshelman, Ed., Morgan kaufmann Publishers,
pp. 73–80.
[6] M ANCORIDIS , S., M ITCHELL , B. S., C HEN , Y.-F., AND
G ANSNER , E. R. Bunch: A clustering tool for the recovery and maintenance of software system structures.
In Proceedings; IEEE International Conference on Software Maintenance (1999), IEEE Computer Society Press,
pp. 50–59.
[7] M ANCORIDIS , S., M ITCHELL , B. S., RORRES , C.,
C HEN , Y.-F., AND G ANSNER , E. R. Using automatic
clustering to produce high-level system organizations of
source code. In International Workshop on Program
Comprehension (IWPC’98) (Ischia, Italy, 1998), IEEE
Computer Society Press, Los Alamitos, California, USA,
pp. 45–53.
[8] M ITCHELL , B. S. A Heuristic Search Approach to Solving the Software Clustering Problem. PhD Thesis, Drexel
University, Philadelphia, PA, Jan. 2002.
8 Acknowledgements
[9] M ITCHELL , B. S., AND M ANCORIDIS , S. Using heuristic search techniques to extract design abstractions from
source code. In GECCO 2002: Proceedings of the Genetic
and Evolutionary Computation Conference (New York, 913 July 2002), Morgan Kaufmann Publishers, pp. 1375–
1382.
We would like to thank Spiros Mancoridis and Brian
Mitchell at Drexel university for their help with software
clustering and Simon Taylor at Brunel university for allowing the work