380
S.G. Robertson . A.K. Morison
models failed to successfully estimate the age of this species. Despite the
promising progress, the results are still below the performance of human readers
who usually achieve APE values below 1% for A. butcheri and P. auratus, and
below 5% for M. novaezelandiae (Morison et al. 1998a). Nevertheless, the PNNs
correctly classified a much larger range of age classes than previous studies have
managed (e.g. Welleman and Storbeck 1995; Troadec et al. 2000). Also, these
results were achieved on a representative subset of otoliths and without the need
for any apriori growth patterns or user intervention.
There are still several issues that would benefit from further investigation.
Some of the apparent errors of the neural network age estimates may be due to
mistakes made the human reader in estimating the age of the sampIes used to train
the networks. Such errors may affect the ability of neural networks to train
effectively or degrade their performance on the production set. This issue could
only be effectively addressed by the use of all known age sampies to train and test
the network. However, such data sets are rare and difficult to obtain for many
commercially exploited fish species.
One further development that is highly desirable is to test the performance of
neural networks against more conventional multi-variate statistical methods. Few
attempts have been made along these lines for age estimation in fish. (Boehlert
1985) used multiple linear regression methods but there are other well-established
non-linear multi-variate statistical methods that could employed. Neural networks
require such benchmarking to truly test the extent to wh ich they represent a
significant advancement in modelling for age estimation in fish or other ecological
issues.
Acknowledgments
This work was supported by grants from the Australian Fisheries Research and
Development Corporation (project numbers 96/136 and 98/105). Fisheries
Victoria provided funds for the collection and ageing of A. butcheri and P.
auratus. The Australian Fisheries Management Authority provided funds for the
collection and ageing of sampies of M. novaezelandiae. The assistance of many
colleagues at the Marine and Freshwater Resources Institute is greatly appreciated,
particularly Corey Green and Kyne Krusic-Golub for estimating ages, staff of the
Bay and Inlet pro gram for sampIe collection, and David Smith for comments on
the draft manuscript. Steve Ward is thanked for his prompt and insightful
suggestions for implementing the neural networks.
References
Beamish RJ, Foumier DA (1981) A method for comparing the precision of a set of age
determinations. Journal of the Fisheries Research Board of Canada 36, 1395-1400
Beamish RI, McFarlane GA (1983) The forgotten requirement for age validation in
fisheries biology. Transactions of the American Fisheries Society(112), 735-43
S.G. Robertson . A.K. Morison
models failed to successfully estimate the age of this species. Despite the
promising progress, the results are still below the performance of human readers
who usually achieve APE values below 1% for A. butcheri and P. auratus, and
below 5% for M. novaezelandiae (Morison et al. 1998a). Nevertheless, the PNNs
correctly classified a much larger range of age classes than previous studies have
managed (e.g. Welleman and Storbeck 1995; Troadec et al. 2000). Also, these
results were achieved on a representative subset of otoliths and without the need
for any apriori growth patterns or user intervention.
There are still several issues that would benefit from further investigation.
Some of the apparent errors of the neural network age estimates may be due to
mistakes made the human reader in estimating the age of the sampIes used to train
the networks. Such errors may affect the ability of neural networks to train
effectively or degrade their performance on the production set. This issue could
only be effectively addressed by the use of all known age sampies to train and test
the network. However, such data sets are rare and difficult to obtain for many
commercially exploited fish species.
One further development that is highly desirable is to test the performance of
neural networks against more conventional multi-variate statistical methods. Few
attempts have been made along these lines for age estimation in fish. (Boehlert
1985) used multiple linear regression methods but there are other well-established
non-linear multi-variate statistical methods that could employed. Neural networks
require such benchmarking to truly test the extent to wh ich they represent a
significant advancement in modelling for age estimation in fish or other ecological
issues.
Acknowledgments
This work was supported by grants from the Australian Fisheries Research and
Development Corporation (project numbers 96/136 and 98/105). Fisheries
Victoria provided funds for the collection and ageing of A. butcheri and P.
auratus. The Australian Fisheries Management Authority provided funds for the
collection and ageing of sampies of M. novaezelandiae. The assistance of many
colleagues at the Marine and Freshwater Resources Institute is greatly appreciated,
particularly Corey Green and Kyne Krusic-Golub for estimating ages, staff of the
Bay and Inlet pro gram for sampIe collection, and David Smith for comments on
the draft manuscript. Steve Ward is thanked for his prompt and insightful
suggestions for implementing the neural networks.
References
Beamish RJ, Foumier DA (1981) A method for comparing the precision of a set of age
determinations. Journal of the Fisheries Research Board of Canada 36, 1395-1400
Beamish RI, McFarlane GA (1983) The forgotten requirement for age validation in
fisheries biology. Transactions of the American Fisheries Society(112), 735-43