NEURO PEX





NEURO PEX is the first commercial software in the world to calculate design of experiments for non-linear models, neural networks, differential equations and implicit models.

The design of experiments for non-linear models’ goals are:

• If the cost of the experiments is low, to define the optimal experiments which maximize the ratio of the model accuracy to the number of experiments;
• If the cost of the experiments is high, to define the optimal experiments which minimizes the number of experiments.

The innovation in Neuro Pex comes from the treatment of the non-linearities of the underlying model that describe the phenomenon you want to study, and from the consideration of all the possible configurations:

• Knowledge based non-linear models described by non-algebraic equations or black box models based on neural networks;
• Real (with noise) experiments, or computer experiments (without noise);
• Ordinary differential equations and implicit models.

The first users of Neuro Pex were the CEA-DAM (Atomic Energy Agency), IFP (French Petroleum Institute), PSA-Peugeot-Citroën and Rhodia, brought together in an industrial partnership. Potential users are every people in charge of the identification of models or design of experiments, namely:

• People who use non-linear knowledge-based models;
• People who use neural-network-based models;
• People in charge of computer simulation
• Expert in design of experiments for linear models.

Concerning knowledge-based models, the results are spectacular: the number of experiments can be reduced down to the number of coefficients of the model. For neural-network-based models, the number of experiments can be reduced down to three times, or sometimes even down to twice the number of coefficients of the model. Much time, as well as money, can thus be saved. In any circumstance, the accuracy of the models is improved.

Neuro Pex includes an executable program, a user guide and a methodological guide which will remind you the main points of the design of experiments theory for non-linear models. The current version of Neuro Pex 2 contains D-optimality and X-optimality algorithms for a non-linear model with a homogeneous or heterogeneous variance, tools to diagnose the quality of the resulting design and the non-linearities measurements in the underlying model.

NETRAL distributes NEURO PEX in two ways:

• Classical licences: this option is interesting if more than 5 designs are calculated in one year, or if the data is confidential. On the other hand, it asks much time of the user to understand the concepts and the software;
• Consulting services: for the user if he only has one important subject to handle and if he doesn’t want to spend more time than necessary on this project.

Please contact us if you are interested in NEURO PEX and if you wish a demonstration.

Discover the Box-Lucas example (knowledge-based model) in our 51 screenshots tutorial.

The Designs of computer experiments Day took place the 24^th of March 2005 in the Institut Henri Poincaré.

© Netral - June 2007