Creating a network topology with Darwin

Question

Hello,
here are some questions concerning the neural net module at Darwin. I would be grateful if anyone can answer some of my questions:
1. sensitivity analysis: How is importance/sensitivity of each independent variable to the model measured ? (E.g. by measuring the %-increase of the train error if specific variable is taken out of the model ? )
2. Darwin provides the user with three types of activation functions: linear, sigmoid and hypertangent. In a paper with the title "Darwin: A scalable integrated System for Data Mining" , May 14 1997, accessible throug the ORACLE web-sites, some more activation functions have been mentioned : Heaviside, sinusoidal etc. The (most recent) release Darwin 3.7, has only three different activation functions. How can different activation functions be added. I would like to know that, because many application examples from the financial sector used the Gaussian or Radial basis (RBF) function to develop models in financial analysis (portfolio management, prediction of financial crisis etc.)
MOST IMPORTANT QUESTIONS CONCERNING DEVELOPMENT OF NETWORK ARCHITECTURE AND NETWORK TRAINING:
-> Please refer again to "Darwin: A Scalable Integrated System for Data Mining
Overview of Darwin's philosophy, architecture, and functionality for 2.x and 3.x releases, 1997" , May 14 1997, accessible throug the ORACLE web-sites on http://datamining.us.oracle.com/
3. p. 7 : The train and cross-validation mode includes an efficient implementation of the cross validation method for choosing the optimale size of network. It takes advantage of the properties of the error surface, and trains multiple networks with only a marginal increase in time needed for training a single network.
As the notion "Cross-Validation" and especially algorithms using "cross-validation"-techniques to develop a network topology have throughout literature completely different definitions and underlying ideas, may you please explain your understanding of cross-validation by clarifying the algorithm for selection of single-hidden-layer neural net as described above.
4. Learning curve analysis: Does exist a bootstrap loop for the evaluation of a learning curve ? How was the learning curve, as it is presented in the paper above, developed with the aid of a bootstrap-loop explained on p. 11
5. Most recent results in research on generalization performance and its relation to the development of an architecture of a neural net, assume/stress again that overtraining is only an effect of a misspecified, too high machine complexity. Therefore, in this case, conventional training&test-mode is not appropriate for traing a learning machine of optimal complexity.
Consequently, the idea of stopped-training is ambiguous. Only a learning machine with a too high complexity, is capable of approximating the intrinsic noise in the data.
The generalization performance of a learning machine of optimal complexity (i.e.no under-, no overfitting) increases asymptotically without turning point. (in this case: test error decreases asymptotically with the number of training iterations)
Own experiences, results, comments ?
Please note : Following article could answer this question from a more theoretical viewpoint, but hasn't been available to me yet:
Wang, C., Venkatesh, S. S. and Judd, J. S. 1994. Optimal Stopping and Effective Machine Complexity in Learning.
Advances in Neural Networks Processing Systems, G. Tesauro, D. S. Touretzky and T. K. Leen Eds. MIT Press.
Vol. 7, p239.
6. Does Darwin provide the user with a visualization tool in order to illustrate the process of learning by showing network weights? If not, how can this be implemented with the help of Darwin.
This has often been defined as "monitoring" in order to show development of network weights and to make some conclusions.
7. Genetic Algorithms: Can you specify the exact values of the parameters used in the genetic algorithm for network training? E.g. the defined value of the probability of mutation. Please note, that different parameter values (e.g. probability of mutation) used in the genetic algorithm have an effect on the performance of the search for a global optimum .
8. Genetic Algorithm : Why does the g
genetic algorithm has a learning rate ? Normally Genetic Algorithm do not have learning rates, please explain this parameter.
Learning rate are only know for gradient descemt algorithms.
9. Key-field wizzard: How is the importance of an independent variable towards the dependent variable measured ? Which is the underlying correlation coefficient ? The Darwin Help explains the two options to calculate the importance: Gini and cost.
10. E.g. why not use IBM intelligent miner for data preparation instead ORACLE Darwin, because Intelligent miner has two very important processes for data preparation in order to clean data from multi-collinearity:
- factor analysis
- principal component analysis (delivers same results)
null

Answer 1

I have a very similar setup and I'd be happy just to be able to get my MacBook to print to the HP6213!
Are there any guidelines out there for setting up to print to a printer connected to a Windoze infected PC?
btw
I suspect you'd need a printer with a LAN card onboard to be able to print to it without using the Windows machine as a print server.
rgds,
Dave