A brief description of NETRAL : Neural Networks, Design of Experiments, Nonlinear Statistics

A brief DESCRIPTION of NETRAL

NETRAL was incorporated as a société anonyme in 1994 by Jean-Luc Ploix, its current chairman and general manager.

NETRAL is specialized:

Since 1994 in the use of neural networks and nonlinear regression models. Neural networks are tools for statististical data processing and for the modelling, simulation and optimisation of phenomena or processes described by nonlinear models. Regression models can be polynomial models, physical models, ordinary differential equations, computer codes. The applications cover all domains where data and nonlinear phenomena can be found: physics, chemistry, biology, mechanics, astronomy, microelectronic, finance and much more. Our softwares are Neuro One, NDK, Neuro Code and Neuro One's modules Neuro Fit and Neuro Shop ;
Since 2003-2004 in the design of experiments for nonlinear models (DoE) like neural networks, physical models, ordinary differential equations and computer codes. The software Neuro Pex, available since July 2004, is the first commercial software in the world on this market niche.
Since 2006 in the accurate estimation of the shape, the quantiles and the Cpk capability indices of non-Gaussian distributions and probability laws like asymetric distributions, fat-tailed distributions or multimodal distributions (new software Neuro Proba-Cpk).

In 2006, NETRAL's commercial offer is the following:

The software Neuro One which is the company's flagship product since 1997. This software includes the following features:

A module for the design of linear and nonlinear models: linear and polynomial models, perceptron neural networks, radial basis function neural networks, Kohonen map neural networks, any nonlinear algebraic model for static training, hybrid models combining neural networks and physical models, time series (ARMA, GARCH, NARMAX, ...), dynamic models;
A module (Neuro Shop) to merge mono-output models into a single multi-output model (for optimisation purpose with Neuro Fit);
A training module for static regression, dynamic regression, static classification and Kohonen maps. Our Levenberg-Marquardt algorithm is accurate to 10 decimal places;
A fast and powerful module (Neuro Fit) for optimisation of mono- and multi-output models: direct mode, reverse mode, compensation mode and desirability function mode;
An automatic generator in Excel and Matlab of code representing neural networks and physical models;

The software Neuro Developer Kit, available since 2001, provides the tools to integrate Neuro One's neural kernel in third parties software or in dedicated applications developped under Windows.
The software Neuro Code, available since 2002, is an automatic generator of C code librairies representing neural networks and physical models. Those librairies, which include the make and the makefile, are ready for use in highly demanding applications like large scientific calculation (memory space, long processing time, 64 bits), industrial control (Schneider Premium, Siemens, Foxboro, Honeywell, Yokogawa, ...) or embedded applications (QNX, OSE, Wind River, ...). Neuro Code's C code is 10 times faster than its counterpart in Windows because it is fully optimised for every model and does not care about the graphics. The code permits some online training (in the proposed operating system) with direct interaction with the data stored on the computer, which can be crucial if the speed or the application integrity are concerned.
The Neuro Pex software, available since 2004, is dedicated to the Design of Experiments (DoE) for nonlinear models, neural networks, ordinary differential equations and computer codes. The current version 2 includes:

D-optimal design and X-optimal design;
The measurement of the model nonlinearities (2nd order analytical derivatives) and some diagnostic tools to evaluate the quality of the proposed design.

The software Neuro Proba-Cpk, available in Autumn 2006, provides an innovative approach based on neural networks to estimate with a high accuracy the shape, the quantiles and the Cpk capability indices of non-Gaussian (non-normal) distributions. Those distributions can be asymetric distributions, fat-tailed distributions or multimodal distributions. The expert version permits some batch calculation up to 400 distributions estimated in one run with a concise summary report.
Some consulting services on the right use of neural networks and nonlinear modelling;
Some advice on the right use of Design of experiments for nonlinear models;
Some technical studies from 2 days up to 6 months on topics defined by the client;
Some training courses on neural networks and Design of experiments. Netral realizes two types of courses:

One five days institutional course (in French) every year in Spring and Autumn in collaboration with the Faculty of Physics and Chemistry of Paris (ESPCI). The courses are in French;
Some courses tailored to the client needs and located in the client premises, usually from one to three days. The courses can be in French or in English;

NETRAL's clients are large industrial groups, resarch and developement centres and university laboratories. We can mention:

Industrial groups: Altadis; Arcelor; Ascométal - Lucchini; CEA - DAM Ile de France (Commissariat à l'Energie Atomique); CF Gomma; Dassault Aviation; EDF; IFP (Institut Français du Pétrole); Nestlé; PSA Peugeot Citroën; PPG (automotive paints); Renault Cars; Renault Trucks; Rhodia; Saint-Gobain; Snecma; SNET Electricité; ST-Microelectronics; Total.
Small companies: Alpha Technologies (Equipment for the rubber industry); Gazonor (Extraction of gas in coal mines); Laboratoire d'Oenologie Dubernet (Analysis of wine);
University Research Laboratories : The centre of Biostatistic and Medical Documentation of Yvoir (Belgium); The laboratory of Chemiometrics of Lyon 1; The laboratory of Atmospheric Chemistry of Paris 12 - Créteil University; The laboratory of Aerology of the Midi-Pyrénées Observatory in Toulouse; The Energetic System and environment Department of Ecole des Mines in Nantes; The faculty of Electrochemistry and Electrometallurgy in Grenoble.
University teaching courses: Faculty of Physics and Chemistry of Paris (ESPCI); Ecole Centrale de Paris; Faculty of Electrochemistry and Electrometallurgy in Grenoble; The Department of Chemistry at University Lyon 1; IUT de Valenciennes.

In fields of activity with little experience in nonlinear statistic, Netral initiates the research, and then transfers the know-how and the results to the client's specialists. This approch gives very good results for a low cost. We often recommand it.

NETRAL's main actors are:

- Jean-Luc Ploix
Jean-Luc Ploix obtained a Master of Science diploma and a PhD in Physics at the Faculty of Physics and Chemistry of Paris (ESPCI: Ecole Supérieure de Physique et de Chimie Industrielles de la Ville de Paris). He served in Thomson and CGR - General Electric on designing and producing medical imaging systems, then in Alstom for the embedded electronic control for the TGV high speed train. He founded NETRAL in 1994 and is the current CEO. He is the main developer of the software Neuro One.

- Patrice Kiener
Patrice Kiener obtained a Master of Science diploma at the University of Technology of Compiègne (UTC). He served in manufacturing departments of Grands Moulins de Paris in France and Yamazaki in Japan then at the industrial direction of Nestlé in China. Since March 2000, he is NETRAL's Marketing and Sales Manager.