|
Roman Ilin
|
 |
Contact
Information : |
Home
Research Area: Computational Intelligence Sub-area: Neural Networks
I study novel types of neural networks able to deal with challenging problems that cannot be solved by feed-forward architectures. Examples of such problems are navigation in complex environments, image processing, playing board games. All of them are cases of intelligent control. They require optimization of the actions over long term
My research focuses on two types of recurrent neural networks. The first type is network with non-convergent dynamics inspired by neural population models called K-sets. I study dynamical properties of these models and also their applications for pattern classifications. The second is Cellular Simultaneous Recurrent Network inspired by the ideas of approximate dynamic programming . I apply this network to 2D Maze navigation problem and Connectedness problem. I use Extended Kalman Filter (EKF) to train this network. Both network types are described in detail if you follow the links.
There is Matlab code for K sets available on CND web site and I also have CSRN Matlab code available here.
The area of Neural Networks is highly interdisciplinary. It has strong links to dynamical systems analysis,numerical methods, neurobiology andneurophysiology , to mention just a few. Such interconnectedness is inevitable in any discipline that has an objective of modeling and understanding a natural phenomenon. Therefore, in the broader sense, my interest lies in modeling of complex non-linear systems, thus covering a broad range of possible areas from brain modeling to economic models to weather prediction.