|
DiscipulusTM genetic-programming software is a powerful regression and classification tool. It has significant advantages over neural-networks, classification-trees, support-vector-machines and robust-regression methods. It is fast, powerful, and has a proven track record of results: Sixty to two-hundred times faster than comparable modeling techniques Consistently out-performs neural networks, classification trees, support vector machines and robust regression methods with less work Unlike neural networks, Discipulus automatically separates relevant and irrelevant inputs and easily handles highly correlated inputs Scaling and preprocessing of inputs is usually unnecessary Intelligent self-configuration means NO parameter setting Creates models from large data-sets quickly and easily Built-in protection against over-fitting Execution in parallel over a Windows network for enterprise level execution speed Easily understood models produced in C, Java or Intel Assembler for deployment
In Genetic Programming, the fitness function evaluates how well evolved programs are solving your problem. The Engineering Version of Discipulus gives you all the power of the Professional Version PLUS complete control over the fitness function. You write a DLL that performs your customized fitness function and then link it to Discipulus, with an easy to use Wizard. Project Automation Automatically sets up each project for you. All you do is tell it where your data is Automatically conducts hundreds of runs and finds the optimal termination Automatically constructs "team" solutions for your modeling problems Automatically finds optimal parameter settings, freeing the you from any need to understand genetic-programming Automatic problem-type detection determines whether you are performing regression or classification and automatically sets Discipulus up for the right kind of run Wizard driven learning makes modeling a simple matter of point and click
Owner's Manual Comprehensive Real-Time Reporting Real-time project overview with project statistics Real-time project progress showing best program and best-team results Real-time Individual run progress and parameters
Comprehensive Post-Run Reporting Best Programs Page gives statistics on the thirty best programs from the project. A simple point and click lets you view and analyze each program or to see its output data-point by data-point Best Teams Page gives you statistics on the five best team-solutions from the project together with voting-based probabilities Input Impacts page provides comprehensive information about which inputs were important and which were not.
Interactive Evaluator Module Import any of the best-evolved programs into interactive evaluator for analysis and simplification, including: View and analyze the thirty best evolved programs in detail Remove unnecessary evolved code (remove introns feature) Optimize the constants in the evolved program Simplify the evolved program automatically Remove, add or change instructions in the evolved program Save evolved programs for future use or deployment Load previously evolved and saved programs from disk
Custom Fitness Functions (Engineering and Enterprise Versions) Two separate interfaces let you write your own custom fitness functions and then link them into Discipulus. Easy to do and very powerful. Parallel Operation over a Windows Network (Enterprise Version) Link five, ten, twenty or more CPU's together on a single run for maximum GP power. Scales almost linearly. All you need is a network of Windows computers and our Enterprise Version. Genetic Programming Operation General Floating Point Operation on x486 and Pentium FPU's Up to 64 Inputs and Constants Configurable number of FPU registers--between 1 and 8 Maximum initial program size and maximum program size. 64 bytes to 2048 bytes
Fitness Function Selectable Function Fitting or Classification learning modes Configurable classification threshold and misclassification penalty Linear or square fitness functions
Functions Available Complete operator set includes arithmetic, conditional jumps, data transfer, exponential, stack operations, and trigonometric operators Choice of operators and the relative proportions of operators are user configurable. Useful preset configurations available. Save and reuse your own custom operator and register configurations
Genetic Programming Parameters Population size--limited only by computer RAM. (Populations of up to 500,000 are easily obtainable on a desktop computer) Demes. Mimic natural evolution by subdividing a population into semi-isolated locales. Crossover and migration allow individual solutions to move between the locales (demes) Dynamic Subset Selection. Eliminates over-fitting problems and improves generalization of solutions. Available for both classification problems and function fitting problems Search Operators. Flexible and user configurable, featuring linear-two-point and homologous crossover, mutation and reproduction Parsimony Pressure. Control how much program size is incorporated into the fitness function to evolve compact solutions
|
