---

Find the fastest BackProp algorithm

We believe that our CEN--Optimisation provides the best training algorithms of the time being. Of course, this claim needs to be verified by using much more benchmarks and the results of CEN--BPROP/RPROP must also be compared to other training algorithms.

In order to do that we propose the following procedure. On the following webpages we provide the results of ten CEN--BPROP/RPROP runs we got with different benchmark problems. If your favorite algorithm can outperform ours, provided you use the same network topology, please create a similar web-page and mail us the URL of that page. In order to get comparable results, we restrict this competition to MFNs with a well defined and fixed network topology. Thus, constructive methods like cascade correlation can't be considered.

If you have investigated a new benchmark problem and you want to know how your algorithm compares to CEN--optimzed ones, create a web-page where the pattern sets are given (as ASCII files), the topology of the MFN is specified and the ten training runs are displayed. Additionally, you can provide a table where the generalization performance (errors/classification rates of the testing set) are given. Mail us the corresponding URL and we will do the same with our CEN--optimized algorithms and publish the results on the web.

The race is on!

Catch our CEN--optimized backprop algorithms!

---

We used two classifikation task from the PROBEN1 benchmarks:

• Thyroid Data
• Gene Data

---

Click here to get a Postscript version of the paper CEN--Optimization (65 KB)
Click here to get a Postscript version (german) of the paper CEN--Optimierung (68 KB)

---

Authors: Merten Joost and Wolfram Schiffmann

last change: 20.3.97

---