Research in neural networks concerns the complexity of learning and computation in neural networks, theory of self-organisation, and embedded neural network tools for data analysis. There are a large number of applications in science, commerce, and industry. Special algorithms are studied for fast and robust self-organizing maps. Software development is made for data analysis.
The interest in complex systems, that is, systems consisting of large numbers of simple components with nonlinear interactions, has been increasing rapidly in recent years because of advances in the theory of nonlinear dynamics, the growth of computer simulation power, and the realization of the ubiquity of such systems in natural and social settings. This project brings together into close collaboration leading research groups in this area from mathematics, physics, ecology, economics, and computer science.
The topic is neural networks and data analysis. Project has been developing a specific neural network model, called a tree structured self-organizing map, that combines projection methods and hierarchical clustering. Very satisfying results have been achieved with fixed default parameters, making the methods suitable for applicants without deeper expertise in neural computing. This method is then used in a larger framework of data analysis software, called NDA (the Neural Data Analysis Environment).
Several application specific products are being developed on the basis of NDA kernel. These are strategical decision making, customer profiling, fault diagnostics, process monitoring, locational market analysis, and data warehousing, for example. Nine enterprices are building their own products, based on the NDA platform.
The aim of this study is to develop computational methods that are widely applicable in human, economic and engineering sciences. We are aspiring after tools for qualitative data analysis to draw valid meaning from both conceptual and exact data. The concrete example is to make a computer aided system to study human learning in technology rich environments, and the methodology is based on a specific neural network model.
We match our methods with a typical problem setting in human sciences, including both cognitive empirical data and structured knowledge of the problem. Data, which are often something else than numbers, are provided by collaborating teams doing research on human learning.
The aim of the project is to adapt some basic deterministic and stochastic methods of dynamic programming onto neural networks (feedback-recurrent and feedforward). The actual goal is to get a powerful tool for handling large-scale optimization problems arising from telecommunication planning. The main topics have been sequential decision processes (like Markov decision problems and other approaches which can be finally set as Markovian decision problems) with a special emphasis on telecommunication network flow control and routing. The theory of diffusion processes and Brownian control are used to be able to describe more accurately the stochastic nature of the arrival and flow processes of the customers in small and medium sized communication networks. The intention is to construct a soft computing tool for planning.
Collaboration with the Telecommunication Laboratory of the Technical Research Centre of Finland.