CAMBRIDGE, Mass. A software package for intelligent operations management from Gensym Corp. helps users with no previous neural-network experience to deploy systems that learn. An aid in complex process control, the NeurOn-Line Studio automatically optimizes embedded deployment while retaining the flexibility, modeling power and efficiency of neural-learning approaches, said Bob Moore, the company's president.
NeurOn-Line includes visualization tools that analyze complex data sets as large as 100,000 records using as many as 100 variables. Models can be analyzed off-line or online in real-time. Off-line, the visualization tools can eyeball historical databases to create accurate models, while online, real-time comparisons with the model enable more-efficient control.
"We are getting substantial value from creating models from data that is just too complicated for engineers to identify on their own," said Moore. "You can find these models, without being an expert in neural networks, to control and optimize your process in real-time."
The Windows-based environment simplifies the creation and management of neural-network modeling and predictions by automating the creation of network architectures, selection of principal components of inputs and fine-tuning of parameters and time constants. Applications are built by selecting appropriate building blocks, arranging their icons on-screen and graphically connecting their input/output nodes.
Four types of building blocks can be wired together graphically to visually program an application. Vector blocks consolidate, condition and manipulate time series and real-time data streams. Data blocks collect, filter, display, summarize and archive incoming data streams. Training blocks prepare data streams with full control over real-time network parameters and validation procedures.
Neural-network blocks learn patterns hidden in data by computing an output vector for each vector of input values. "NeurOn-Line finds the five best neural-network models from among as many as 60 or 70 that it tries out, each model using a different architecture," Moore said.
Applications developed with the NeurOn-Line Studio may be deployed either as ActiveX controls, for embedding in other user-written programs, or as modules for Gensym's G2 environment for intelligent operations management. Deployed applications retain their predictive-modeling and optimization capabilities, regardless of which method is used.
ActiveX deployment simplifies the integration of neural-learning capabilities into existing real-time distributed architectures. As such, the NeurOn-Line Studio can address a wider range of applications than Gensym's traditional process-control customers.