Research Themes

Below are brief descriptions of various research themes that I have worked on, latest work first. Given how research works, there is usually a time-lag between the topics described here; and their fruition as research papers on the publications page.

Knowledge and Model Acquisition for Complex Reasoning

Machine Learning (ML) techniques rule the roost in AI research today, particularly Deep Learning and Reinforcement Learning. While recent advances have shown a focus on learning from scratch, there is still a big role for knowledge and models to play in this brave new world. Specifically, the AI agents and systems of the future must be able to combine learning and reasoning skills to make the most sensible use of all available data as well as prior knowledge. One of my current and immediate interests is in looking at abstract representations of reasoning models (that will accommodate multiple kinds of reasoning techniques) and how to learn these abstract representations from data and exploration; as well as incorporating prior knowledge into such representations. The prior knowledge may either be elicited actively from human experts, or learned passively from specific traces of the agent or system's execution in the world. I am currently working on evaluating this idea on a specific application (Reasoning for Complex Question Answering) and a specific technology (Automated Planning).

Human-Agent Collaboration

My graduate research work was strongly centered around the theme of reasoning and decision-making systems as mediators for human-machine and human-agent teams. This idea encapsulates two distinct technical challenges: first, reasoning formalisms must be made expressive enough to be able to represent real-world scenarios. Then the applications of interest must be modeled in order to use state-of-the-art reasoning systems that have been augmented with these extensions. My research statement elucidates on this idea.

At IBM, I had a chance to apply this to something different from the robotics domain in 2017 -- my work in that period focused on combining symbolic AI techniques like planning with HCI tools and technologies to produce smart interfaces for business applications.

Data-Driven Dialog

A couple of years ago, I worked on the problem of conversational systems that engage in full end-to-end dialog (as distinct from single-turn question-answering). In particular, the work I did focused on the following additional conditions: (a) goal-driven dialog; and (b) data-driven (learned) models. The idea centered around building fully-encapsulated dialog agents that can operate with learned models of a domain to steer the conversation towards the resolution of problems. We evaluated our systems and agents in the context of technical support for Ubuntu (please see publications for our latest reports).

Planning for Human-Robot Teaming

My Ph.D. thesis work centered around the problem of planning for human-robot teaming; that is, considering the automated planning issues involved when supporting the execution of a robot in cooperative scenarios that involve teaming up with humans. Specifically, the problem was to have a high-level task-planner control -- and plan for -- a robot that has to act in a constantly changing world.



Google Scholar

Here is a link to my Google Scholar page, which has a handy (and current) citation count.

Erdös Number

My Erdös Number is 3: William Cushing ⇄ Henry Kierstead ⇄ Paul Erdös


IBM TJ Watson Research Lab

In 2012, I interned at IBM's TJ Watson Research Laboratory in Hawthorne, NY. My work was related to streams, and automating the process of composing them and handling the entities and artifacts that are produced as a result of these compositions. We handled a network traffic monitoring application, and used planning tools and models to refine the security alerts that are escalated to a human administator, thus improving overall efficiency and reducing the response time to serious incidents.

IBM India Research Lab

I worked with IBM's India Research Lab in Bangalore (in 2011) on creating a model for road traffic and various intelligent sensors and systems that are part of the 'Internet of Things' (IoT). This model was used with planning systems to produce automated decisions on controlling the traffic in order to optimize metrics like throughput and wait time.


Refereeing (Selected)