Tarun Gupta – Making the Right Decisions

The Association for the Advancement of Artificial Intelligence (AAAI) recently concluded its annual conference in New Orleans, USA. According to it’s website, “The purpose of the AAAI conference series is to promote research in artificial intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers in AI and its affiliated disciplines. AAAI-18 is the Thirty-Second AAAI Conference on Artificial Intelligence.”

IIIT-H student Tarun Gupta’s paper on “Planning and Learning for Decentralized MDPs with Event-Driven Rewards”, competed against over 3800 submissions and was accepted for an oral presentation at the conference.

Novel Algorithms

“An MDP or Markov Decision Process is very much like a decision making problem, i.e. when you are confronted with a decision, you have a number of alternatives to choose from. When you select an action from one of the given alternatives, you are given a feedback indicating whether your action revealed a positive (good) or negative reward on yourself. Overall, we try to maximize the long term reward.”, explains Tarun. MDPs find applications in a lot of areas including robotics, sensor networks, and general decision making problems.

To explain Tarun’s research work more simplistically, let’s take the case of a driverless car. It is an ‘agent’ that has to work independently and take decisions but in an environment consisting of multiple other agents (perhaps other cars driven by humans or more driverless cars) in order to function effectively. “Our work formulates policies for such agents to operate in a decentralized manner (no communication, no interaction with other agents) and yet, maximize the overall coordination and make the task successful. Existing algorithms find it difficult to solve such policy formulation problems for a large number of agents (>50)…Therefore our paper contributes 3 novel algorithms for solving Decentralized-MDPs with improved scalability. The Deep-Reinforcement Learning based algorithm is able to scale up to a real-world sized problem and gives great results, “ he claims.

Lateral Entry Candidate

Tarun is a lateral entry final year Dual Degree (B.Tech & Master of Science in Computer Science and Engineering by Research) student who transferred to this program in the second year from Dhirubhai Ambani Institute of Information and Communication Technology (DAIICT). His love for rational problem-solving has been the force behind the strong interest in mathematics he has demonstrated throughout his schooling. He claims to be passionate about computer programming as a tool to simplify and expedite problem-solving. Reiterating a commonly known fact about the dual degree program at IIITH, that it provides a stimulating environment for academic inquiry, Tarun says he would recommend it wholeheartedly to anyone who is keen on research.

Academic Credentials

Tarun has been guided in his research area by his professor Dr. Praveen Paruchuri from IIITH and co-advised by Dr. Akshat Kumar from Singapore Management University (SMU). This collaboration with SMU started in 2016, when I was selected and sponsored to attend the summer school organized by AAMAS-2016 (one of the biggest A* conference in multi-agent systems) in Singapore. I was introduced to the faculty at SMU (Akshat Kumar), who is now my co-advisor.” Graduating at the top of his batch and earning a gold medal in the process (which he will receive during the convocation in August 2018), Tarun plans on pursuing his interest in AI and ML next year . “From mining data on previous movies watched to recommend additional movies (for the viewer), to auto collision prevention in driverless cars, to machines surpassing the human efficiency in recognizing diseases – I believe AI and ML technology has the potential to change the world by transforming the ways in which knowledge can improve people’s lives, “ he says.

Sarita Chebbi is a minimalist runner, practising yogi and baker of all things whole-wheat, and sugar-free. Currently re-learning her ABC’s…the one that goes: A for algorithm, B for Bayesian, C for convolutional (neural network)….

Leave a Reply

Your email address will not be published. Required fields are marked *

Next post