MIT EECS - VMware Undergraduate Research and Innovation Scholar
Building Elasticity Module for FlexGP
Flex-GP is a large scale, cloud-based machine learning system that runs genetic programming. The system has an island-based distributed model that communicates through socket-based client-server protocol. The cloud platform EC2 from Amazon was primarily used for the implementation and evaluation of the algorithm. The systems performance was evaluated on a regression problem with various numbers of nodes, ranging from 3 to 350 nodes. The accuracy, which is the best mean squared error (MSE) achieved on the regression problem, improved significantly as the number of nodes increased, achieving the best average MSE for 256 nodes. The compute time the overall time to achieve as high level of accuracy as the computation on multiple nodesincreased by a factor of 3 when the 255 nodes were launched in addition. The performance of Flex-GP can be further explored as it runs on more nodes, employing various additional techniques like ensemble based machine learning.
I have worked in Learning Interlligent Systems group, developing application for PR-2 robot. I did an internship at Vecna Technologies building an independent ROS node that detects walls in a hallway by filters out noise.