After the skills of our model developers, PreMo’s belief that it can produce more accurate models than others is based on the capabilities of our favorite tool, Discipulus 4.0.  Discipulus is a genetic programming predictive modeler and has become the ‘go to’ tool of choice at PreMo.  All of our models are built using at least two different tools and in regression or classification, one of those will be Discipulus.

Genetic programs operate by starting with a series of randomly generated algorithms based on your data, then running series of competitions among them to see which results in the best answer.  While this is happening, genetic programs use Darwinian or evolutionary logic to mutate the programs and to allow different competing algorithms to swap elements similar to the way viruses might swap genes.

There are several keys to success but the most obvious is that genetic programs must be exceedingly fast to run competitions among the millions or tens of millions of variants in a reasonable period of time to discover the best answer.  While increases in computing speed have allowed good performance for many genetic programs, Discipulus is the only genetic program written at the machine register code level.  Its patent protected engine executes calculations at the chip level.  No compiled code is used for the calculations which are performed directly within the FPU registers of any WINTEL chip.  What this means is that it can run 100 to 1000 times faster than others.  Speed means accuracy.  With the ability to examine 100 to 1000 times as many potential solutions in the same period of time, Discipulus is 100 to 1000 times more likely to find a better solution than others, and regularly does.

In addition, Discipulus has a number of other features that guarantee higher accuracy:

Automated Tuning Parameters: Only Discipulus 4.0 has fully automated the extremely complex and time consuming tasks of tuning the program’s control parameters.  This assures that parameters are optimally tuned and allows unparalleled levels of accuracy.

Single Solution or Team Solution: Through a sophisticated routine separate from determining the best single program, Discipulus evaluates and caches the best programs which working together, reveal the optimal answer.  Discipulus forms teams of 3, 5, 7, and 9 separate programs and periodically develops a control (voting) logic so that these teams of programs can predict or classify the data.  Frequently this team answer is superior to any single model.

Detects the Weakest Relationships:  A combination of raw speed and its internal design for genetic mutation and crossover results in Discipulus detecting and utilizing even the weakest relationships in data.  This capability results in Discipulus solving problems that no other predictive modelers had previously been capable.

Minimal Data Preparation:  Unlike some other categories of programs, Discipulus does not require data normalization.  This can represent a time savings of several days labor on some problems.

Discipulus 4.0 takes on the leading decision trees, neural nets, and advanced regression tools and beats or ties them in every test, setting a world-class record for one industry benchmark data set.(white paper)

Interested in more detail?  Read this technical whitepaper describing Discipulus’ unique approach to genetic programming.