The topics below are guaranteed funding and interest. That said, we are open to other topics!
LLM Agents and Robots that Reason, Plan, Act, and Learn
[Related: Cognition, AI, NLP, Planning, Robots]
The phenomenal success of applications using LLMs for question answering leads to challenges in using natural and artificial languages to describe knowledge, as a basis for reasoning and inference. We are very interested in investigating how LLMs can be used in the construction of intelligent agents and robots.
Recognizing Goals and Plans Interactively
[Related: Cognition, Plan- and Goal-Recognition]
We are developing the next generation of algorithms for
plan recognition (inferring someone’s plan from observations of their actions),
and goal recognition (inferring their goal).
There are several open challenges in this area that we think are particularly interesting:
- Novel goal recognition: Current algorithms assume that the set of goals (and their priors) are fixed, but humans seem to recognize goals dynamically. How can we recognize goals unknown to us?
- Epistemic goal recognition: If we someone looking through drawers, we understand they are looking for something. Their goal is one of knowledge. Yet almost all existing research ignores such goals.
- Continual, life-long, plan and goal recognition. Goal recognition has always been treated as a one-shot problem solving. But humans carry it out all the time, in parallel to their activities. How can this be achieved?
Computational Models of Nanorobot Swarms
[Related: Theory (distributed systems)]
Molecular robots are the future of medicine.
They work as swarms—trillions of small machines—each a small finite-state machine, at best.
By interacting with each other locally, they can cause state changes, and thus carry out a
distributed computation. Employing population protocol models,
we would like to build algorithms and provide guarantees on what can be computed, how well
it can be computed, and how much time it will take. These theoretical questions have
important implications for future medical applications.
Reinforcement Learning, and Rational Swarms
[Related: Reinforcement Learning, Robots, Game Theory, Cognition]
Using reinforcement learning in robot swarms, our robots learn to coordinate with each other, despite being only capable of local interactions. This works really well in some tasks, but in others… Not so much. We are investigating new algorithms and methods for reinforcement learning in robot swarms. Some example challenges include:
- Theoretically, when will difference rewards (the reward function we use) succeed in swarms? Where will it fail?
- How can reinforcement learning be used in swarms, where there is a large temporal difference between the time the actions are taken, and the reward provided?
- Can the rational swarm model be applied to formation flying? herding, schooling and other forms of collective motion?
Predicting the Future: Social Simulation with Qualitative Reasoning
[Related: AI, social and political sciences]
The goal of the thesis would be to explore the use of qualitative modeling methods (an established area of AI) to model and predict social phenomena. Qualitative modeling allows scientists to represent the qualitative factors that influence the system, and their interactions, without using exact measures (which they don’t have). Fridman and Kaminka (TIST 2013) showed that this can be done to predict the level of violence in demonstrations.
This is a good topic for someone with political science, social sciences, or macro-level economics background. It would involve: finding what to model (e.g., peace negotiations? economic crises? urbanization?), modeling using existing tools, and then extending the tools with new algorithms. It does not require a lot of programming or heavy computer science background.