BMS Creative

Results of prediction

N-Expert employs the technology of feed-forward artificial neural networks, trained in several stages by means of the improved back-propagation algorithm. Some new training parameters have been introduced and they are computed in real time, thus allowing one to control the entire training process. Contrary to some other software, N-Expert doesn't offer a wide variety of network architecture, training methods etc. However, this is not a disadvantage, since it is extremely difficult for humans to adjust such free parameters. N-Expert does it automatically using a special algorithm RDO (Reduced Data Optimization), developed by our company and well proven in practical applications.

Scheme of the neural networks training process

Since the training of neural networks is usually a time-consuming process and for some complex problems it may take even hundreds of hours, the efficient use of neural software needs often very fast computers, what of course implies high cost. For this reason N-Expert employs a special technology of distributing computations over usual PC's in a local area network (grid computing). As shown on the picture below, this is done by a supervising module communicating directly with an Oracle database and with so-called NNA clients (Neural Network Advisors), installed on particular workstations. They are given their jobs to do and they are permanently controlled by the supervising module. The communication is based on the D-COM technology.

System architecture

The described method of distributing computations has already been employed by our company in the portfolio management supporting system PORSENNA DSS. Below is the data processing scheme, showing the possible sequences of operations performed by N-Expert.

 Data processing scheme

Learn more about neural networks.