Hi people, libOptimization 1.0 build 18.12.00 has been released. You can grab it at http://svedaq.tsl.uu.se/~anton/neural.html the lib includes: * Simulated annealing * Genetic algorithm (interface to GAlib from MIT) * Coordinate descent It can be used to minimize TF1 and TF2 or any object derived from TOptimizable class. The package includes two examples on TF1 and TF2 minimization (sin(x)/x alike). Simulated annealing algorithm is modified to easily solve different combinatorial optimization problems, for example to find the best subset of N items from M item universe according to a (non-quadratic) objective (cost) function. Examples of real-word applications, such as portfolio/trading strategy optimization can be found on the R-Quant web site. The algorithm has shown quite good performance selecting the best 40 stock portfolio from 1.500 stocks presented on the market in a reasonable time of a few minutes/tens of minutes (depending on a desired error tolerance). The objective function of this 1.500 parametric problem with C(N,M) local minima included step-wise/ non-quadratic transaction costs. On the R-Quant web site you can also find some theory and references about simulated annealing and its modifications. Note : If you compile libNeural with libOptimization, you can train nets with the algorithms above. Although stochastic training is usually (much) slower than conventional gradient propagational methods, it guarantees that the net ends up in the global minimum of its error function. Thus, for example, a net can be trained to solve spiral separation problem (two classes of points on 2D plane separated with a spiral border). This can be in no way achieved with backprop or its modifications. libOptimization is a part of R-Quant Data Analysis Studio. Several screenshots of a new Qt based IDE is available on the site. libNeural and libOptimization classes can not be used in any commercial project or non-scientific environment without written permission ganted by the author. If you use these classes in the data analysis, etc., please let me know to stimulate further development in right direction. /Anton
This archive was generated by hypermail 2b29 : Tue Jan 02 2001 - 11:50:39 MET