Who am I?: I am Sri Ramana. I completed my Bachelor of Engineering in Electronics and Instrumentation in 2014 at Madras Institute of Technology(MIT), Anna University, Chennai. I am currently a Project Associate under Dr. Balaraman Ravindran.
Areas of Interest: Reinforcement Learning, Robotics, Machine Learning
Areas currently working in: Reinforcement Learning
Research Topic: Instruction framework for Bayesian POMDPs
We are looking at ways to incorporate human instructions in a bayesian PODMP setting where the dynamics and rewards are unknown, so as to increase rate of learning. One of the main focus of instruction framework is to design them simple enough so as to reduce the cognitive load on the human instructor at the same time enabling framework to make use of the information in the most effective manner possible. Currently we are looking at two specific kinds of instruction, the 'Pi' and 'Keyword' instructions. In 'Pi' instruction the human instructs an agent about the optimal action to take at a particular state. The framework makes use of that instruction to increase the model's belief. To motivate the 'Keyword' instruction, imagine the agent is in a particular state and the human observes and wants to reveal to the current state to the agent. He says "You are near the table", here "near the table" is the keyword. The problem is that agent doesn't know what it means. But after repeated occurrences of "you are near the table" keyword instruction, the agent actually learns what it means and it effectively uses this newly learnt information.