|
Ameet's World
|
|
|
Current Work (PhD)I am working with Dr. Lalita Udpa in Non Destructive Evaluation Lab (NDEL) at MSU. My work includes non destructive testing of gas pipelines using Magnetic Flux Leakage (MFL) technique. The work includes the noise removal in the raw MFL data. New methods are developed to tackle this problem: (1) state of the art Proportionate Normalized Least Mean Square (PNLMS) techniques. (2) Local Kernel based filtering using the higher order statistics, (Skewness and Kurtosis). Previous Work (MS)I worked with Dr. Juyang Weng in Embodied Intelligence (EI) lab at MSU. The focus of the work was "Autonomous Mental Development (AMD)" of robots. We worked on robot learning in high dimensional and continuous state and action spaces in the AMD mode. I developed two new methods based on the reinforcement learning paradigm, DUDC and HCGS. DUDC stands for Direct Update on Direction Cosines and HCGS stands for High Dimensional Conjugate Gradient Search. The methods were tested on a challenging application of autonomous learning of speech by robots. The target was to teach the robot five basic vowels, /a/ as in car, /i/ as in bit, /u/ as in do, /e/ as in bet and /ae/ as in bat. The novel part in this learning was nothing related to human speech was programmed in the algorithm. Using a generic algorithm based on the two proposed methods, using purely non-supervised reinforcement learning the vowels were taught. Here is a video demonstrating a unique experience of interactive learning of robots to learn to speak basic vowels. The learning program is written based on the DUDC method of learning in high dimensional state and action spaces. The video file is about 35MB and a broadband connection would be necessary to view it online. Publications
|